# 5 Best Memory Tools for AI Coding Agents (2026) Summary: Memory tools for AI coding agents, taken from vendor pages, for a builder whose agent forgets the last session. Five options. Checked 3 Nov 2026. Canonical: https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents Updated: 2026-11-13 Author: Dharmik Jagodana Tags: [#AI Agents](https://www.dharmikjagodana.com/blogs/tags/ai-agents.md), [#coding-agents](https://www.dharmikjagodana.com/blogs/tags/coding-agents.md), [#tools](https://www.dharmikjagodana.com/blogs/tags/tools.md) ## Key takeaways - Memory tools for AI coding agents keep what a coding session already did, so the next session can use it. - First is Beacon, a local session record. Setup preselects Beacon Cloud unless you opt out. - Beacon keeps a session record. claude-mem writes notes for the next session. Basic Memory is Markdown. Mem0 and Supermemory are APIs. - Local Beacon, the claude-mem engine, and Basic Memory's open source are free. Mem0 starts at $0. Supermemory's free plan includes $5 of credits. Memory tools for AI coding agents are add-ons that keep what a coding session already did, so the next session can use it. The five below are ordered by how closely each vendor describes that job. That order is not a score from using them. [Beacon](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#beacon) keeps a local record of the session and can forward it. [claude-mem](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#claude-mem) turns a session into notes the next one can read. [Basic Memory](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#basic-memory) is a Markdown knowledge base you and the agent both edit. [Mem0](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#mem0) is a memory API, with coding-agent plugins that call its platform. [Supermemory](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#supermemory) is a hosted memory API, with a local binary that is not the public source. The notes below are vendor pages, as of 3 Nov 2026. I have not run these products. No link in this post is an affiliate link. - Best for a local session record, with cloud as an opt-in: [Beacon](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#beacon). - Best for notes the next session can read: [claude-mem](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#claude-mem). - Best for Markdown you and the agent both edit: [Basic Memory](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#basic-memory). - Best for a memory API the coding plugin calls: [Mem0](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#mem0). - Best for a hosted memory API with a separate local binary: [Supermemory](https://www.dharmikjagodana.com/blogs/best-memory-tools-for-ai-coding-agents.md#supermemory). Updated November 3, 2026. - Scheduled for 13 November 2026. Vendor pages checked 3 November 2026. No hands-on testing. - Re-check prices and facts by early December 2026. Re-open Beacon's Team price, Mem0's usage rate, and the Basic Memory pricing page. ## How this list was chosen An assistant drafted this from the vendors' own pages and repos. I have not used Beacon, claude-mem, Basic Memory, Mem0, or Supermemory. There is no hands-on test in this post. Checked on 3 November 2026. Each tool section links the page the sentence comes from. When a page is silent, the section says "not published". The comparison table says "not stated" for that same silence. When two of a vendor's own pages disagree, both are quoted. The order is relevance to coding-agent memory, as the vendors describe it. A product whose own page is about remembering a coding session comes before a general memory API. It is not a score from using them, and it is not one winner for every reader. The five were picked on these criteria, in this order: - The vendor's own page describes memory for a coding agent you already run, not a new agent that replaces it. - The page was open on 3 Nov 2026. A third-party roundup was not a source. - The set was already these five. Nothing was added to round the count, and nothing was dropped to change it. - A stated best-for line comes from what the page says the product stores, not from a score. Affiliate status did not change who is on the list or the order. No link in this post is an affiliate link. I have not joined a program for any of these. None of my products is in the list. What did not fit is in the section after the five. [agentmash](https://agentmash.dev/) is named there only to point at the [multi-agent workspaces post](https://www.dharmikjagodana.com/blogs/best-multi-agent-coding-workspaces-for-solo-builders.md). It is not a memory tool. ## Comparison The tool name stays in the first column while the other columns scroll. Comparison of Memory Tools for AI Coding Agents, vendor data as of November 3, 2026 | Tool | Best for | Starting price | Free tier | MCP support | AI models | Platforms | Open source/self-host | Standout feature | Main drawback | Runs locally | Coding-agent plugin | Data location | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | [Beacon](https://beacon.sh/) | A local record of coding-agent sessions, with cloud forwarding as an opt-in | Pro $20 a month, billed annually | Local is free. Beacon Cloud is $0 with 50 findings | Local server, and an optional cloud server | Beacon Cloud uses OpenAI and Anthropic. A local model is not stated | macOS, Linux, Windows, CI, and a browser extension | MIT CLI. Beacon Cloud is proprietary | One event model across many coding agents, stored as local JSONL | Setup preselects Beacon Cloud. Cloud research use is on unless you opt out | Yes. Default is local JSONL | Hooks or plugins for 29 local runtimes | Local by default. Cloud region is not published. Enterprise can use its own bucket | | [claude-mem](https://cmem.ai/) | Session notes a later coding session can read, with an optional hosted observer | CMEM Pro $30 a month | The engine is free. CMEM Pro has a 30-day trial | CMEM Pro gives a private MCP link | Hosted observer, your OpenRouter or Gemini key, or your Anthropic plan | Claude Code plugin and installers for other coding agents | Apache-2.0 engine. CMEM Pro is commercial | Compresses a session into notes and injects them later | The default installer signs you into the hosted observer | Yes. SQLite and Chroma on the machine | Claude Code, OpenCode, T3 Code, Pi, and others | Local unless cloud sync is on. The default installer still uses the hosted observer. Region not published | | [Basic Memory](https://basicmemory.com/) | A Markdown knowledge base you and the coding agent both edit | Cloud $19 a month per workspace | Open source and self-host are free. Cloud trial is 7 days | Official MCP, stdio or https | not stated | MCP clients, Obsidian, and a cloud workspace | AGPL-3.0 | The knowledge is Markdown files, not only a vendor database | The README still lists a cloud price the pricing page no longer sells | Yes. Markdown files and a local index | Claude Code plugin, plus Hermes and OpenClaw | Local files, or per-tenant Tigris and Neon. Region not published | | [Mem0](https://mem0.ai/) | A memory API for app builders, with coding-agent plugins that call the platform | Free ($0), then usage billing | Free tier. Mem0 pays model calls for 30 days, within published limits | Hosted MCP, plus a local search tool in the Claude Code plugin | Your LLM key, or Mem0's for the first 30 days. Model names are not stated | Plugins, a hosted MCP, and a self-hosted library | Apache-2.0 library and server. The platform is proprietary | OpenMemory is the coding-agent layer on the same platform | The coding plugins require a platform API key. The per-call rate is not published | The OSS server can. The coding plugins call Mem0's platform | Claude Code, Cursor, Codex, OpenCode, and others | Platform servers. Region not published | | [Supermemory](https://supermemory.ai/) | A hosted memory API, with a local binary that is not the public source | Pro $19 a month | Free plan with $5 of credits a month | Hosted MCP with OAuth | Local mode can use Ollama. Cloud models are not stated | Plugins, hosted MCP, and a local API on port 6767 | Repo and plugins are MIT. The local server binary is not public source | A free local binary up to 10,000 documents, separate from the MIT repo | The docs also say self-hosted deployments are a Scale and Enterprise option | Yes, as Supermemory local, up to 10,000 documents | Claude Code, Cursor, Codex, OpenCode, Muse Code, and others | Hosted by default. The policy says data may be processed in other jurisdictions. Region not published | ## Beacon [Beacon](https://beacon.sh/), from Asymptote Labs, is an open-source memory layer for coding agents. The site's tagline, as of 3 Nov 2026, is "Self-improving memory for coding agents." It captures session history and turns workflows, corrections, and debugging patterns into knowledge an agent can get back through MCP or Agent Skills. The main limit is the setup: the interactive install preselects Beacon Cloud, and you have to opt out to stay local. The [README](https://github.com/asymptote-labs/agent-beacon) and the site call it a memory layer. Investors and the press call the same company an open-source telemetry layer and an agent-security data layer. The README lists Claude Code, Cursor, Codex, OpenCode, and 20 or more other agent runtimes. Its table names 29 local runtimes, including Antigravity CLI, Claude Code, Claude Cowork, Cline, Codex CLI, Codex Desktop, Cursor, `DeepSeek Harness`, Devin CLI, Devin Desktop, Factory Droid, Gemini CLI, GitHub Copilot CLI, goose, Grok Build, Hermes, Kimi Code, Kiro, OpenClaw, OpenCode, OpenHands, Pi, Qwen Code, and VS Code. Claude.ai and ChatGPT connect through a browser extension. Claude Code, Cursor, and Devin also have cloud agents. CI uses `beacon ci exec`. SDKs cover Anthropic, OpenAI, the Claude Agent SDK, and the Vercel AI SDK. Capture depends on the runtime: hooks, OTLP, plugins, or polling. The event model is OpenTelemetry. The default store is local JSONL at `~/.beacon/endpoint/logs/runtime.jsonl`, with a local terminal browser and a read-only local dashboard. For that local use, the project says nothing is sent anywhere. Beacon Cloud is optional hosted forwarding. The interactive setup preselects it. Local is an explicit opt-out. Capture has two modes, Standard and Metadata-only. There are two MCP servers. `beacon mcp serve` is local and never touches the network. `beacon-managed` is the cloud MCP. `beacon mcp connect` registers it, with OAuth, in Claude Code, Codex CLI, Cursor, VS Code, Gemini CLI, and OpenCode. Skills install to `~/.agents/skills` and `~/.claude/skills`. You can also forward events to Splunk, Datadog, Elastic, Sentinel, CrowdStrike LogScale, Sumo Logic, Wazuh, S3, GCS, and CloudWatch. The CLI is MIT. Beacon Cloud and the enterprise platform are proprietary hosted services. The [pricing page](https://beacon.sh/pricing), as of 3 Nov 2026, shows no date on the page. - Beacon Cloud: $0, described as forever, with 50 findings. - Pro: $20 a month, billed annually. The page says "Save 25%". A monthly Pro price is not shown. - Team, on the card: $30 per seat a month plus a $10 a month base fee, up to 20 seats, billed annually. - Team, on the compare table: $40 per seat a month. - Enterprise: custom. - Local: free. Beacon's pricing card says Team is billed annually; the compare table shows $40 per seat a month. macOS setup, from the project: ```bash brew trust asymptote-labs/tap && brew tap asymptote-labs/tap && brew install beacon && beacon endpoint install ``` Linux is an install script, or a `.deb` or `.rpm`. Windows is an MSI. Or tell the agent to read [the setup file](https://beacon.sh/start.md) and set Beacon up. Claude Code can add the plugin with `/plugin marketplace add asymptote-labs/agent-beacon`. The [privacy page](https://beacon.sh/privacy) is effective 15 Oct 2026. A local or self-hosted destination sends nothing to them. For Beacon Cloud, "Research Uses" are on unless you opt out. That covers training and fine-tuning models, staff review, and selling or licensing de-identified datasets to third parties. An opt-out in Settings takes effect within 30 days. Models already trained are not retrained. Cloud features use OpenAI and Anthropic under zero-data-retention terms. Detected secrets are excluded from research datasets. Enterprise can use its own bucket and region. The standard cloud region is not published. Latest release, as of 3 Nov 2026: v1.3.52 on 2 Nov 2026 IST. GitHub stars: 3,180. The repo was created 12 May 2026. On [Product Hunt](https://www.producthunt.com/leaderboard/daily/2026/11/2), Beacon was the top product on Product Hunt on 2 Nov 2026. [Asymptote's post](https://www.asymptotelabs.ai/blog/asymptote-labs-raises-5-5m-seed-round), [Tower Research](https://tower-research.com/asymptote-labs-the-system-of-record-for-agentic-work/), and [SecurityWeek](https://www.securityweek.com/asymptote-labs-lands-5-5-million-seed-for-ai-agent-runtime-security/) say the company announced a $5.5M oversubscribed seed on 20 Oct 2026, led by Tower Research Ventures, with Lytical Ventures, 1984 Ventures, and South Park Commons. The founders named there are Justin D'Souza (CEO) and Shukan Shah (CTO). The repo has existed since May 2026. The seed date is the announcement, not the day the project appeared. No public affiliate program was found. beacon.sh/partners and beacon.sh/affiliates returned 404 on 3 Nov 2026. That did not change this row. Use it when you want one local log of many coding agents, and you are willing to opt out of the cloud during setup. Skip it when you need the cloud region in writing before any session leaves the machine, or when you do not want research use of cloud data turned on by default. ## claude-mem [claude-mem](https://cmem.ai/), by Alex Newman (@thedotmack), is an open-source memory engine. It captures what the agent does in a session, compresses that with AI, and injects relevant context into a later session as structured notes: decisions, discoveries, and fixes. CMEM Pro is the paid hosted layer: hosted observation, so memory generation is off your own agent plan, plus cloud sync across machines and a private MCP link. The main limit is the default install. It asks you to sign in, and that sign-in uses the hosted observer. The [integrations page](https://cmem.ai/integrations), as of 3 Nov 2026, lists automatic capture through installers or plugins for Claude Code (the plugin marketplace, or `npx claude-mem install`), OpenCode 1.3.4 or later, T3 Code, Antigravity CLI, Oh My Pi, Pi, `DeepSeek Harness`, OpenClaw, and Grok Bot by watching the chat log. The GitHub description also lists Codex, Gemini, Hermes, and Copilot. Recall only, through the CMEM Pro MCP link, covers ChatGPT (read-only, and it does not capture ChatGPT chats), Cursor, Codex, Gemini CLI, Windsurf, and VS Code extensions. Storage is lifecycle hooks: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, and SessionEnd. A local worker uses SQLite with FTS5, and Chroma for hybrid search. Settings live in `~/.claude-mem/settings.json`. `` tags keep content out. Screenshots are opt-in. You pick who generates the memory: the hosted observer, your own OpenRouter or Gemini key, or your Anthropic plan. The engine is Apache-2.0. The [README](https://github.com/thedotmack/claude-mem) points at the [docs](https://docs.claude-mem.ai/introduction) for the open and commercial boundary. CMEM Pro, cloud sync, and TeamBrain are commercial. The [pricing page](https://cmem.ai/pricing), as of 3 Nov 2026, shows no date on the page. - Engine: free. Model or provider usage may still cost extra. - CMEM Pro: $30 a month after a 30-day free trial. A card is required. Prepaid cards are not accepted. Billing can start when the trial ends or when the included usage is used up, whichever comes first. - Heavy: $70 a month. - CMEM Max: $100 a month, plus billed overage above the allowance, "subject to a safety ceiling." - TeamBrain: a pilot only, with no public price. Auto-upgrade settings can move you up. The allowances are not published. The page says they are shown at checkout. Install: ```bash npx claude-mem install ``` Or, in Claude Code, `/plugin marketplace add thedotmack/claude-mem` and then `/plugin install claude-mem`. Other agents use `--ide` with one of `opencode`, `t3code`, `antigravity`, `omp`, `pi`, `dsh`, or `grok-bot`. The install needs Node, plus Bun and uv, which it installs for you. `npm install -g claude-mem` installs only the SDK. The [privacy page](https://cmem.ai/privacy) says "Local-first by default... Nothing leaves your device unless you explicitly turn on cloud sync", and "we do not sell your data, and we do not use the content of your memory to train models." The README says the installer asks you to sign in by default, which turns on the hosted observer and a 30-day free trial of hosted memory generation. When the trial ends, memory falls back to your Anthropic plan unless you subscribe. In the default flow, session content may go to the hosted observer even when cloud sync is off. The privacy page does not address that. The region is not published. Latest release, as of 3 Nov 2026: v15.9.0 on 2 Nov 2026. GitHub stars: 111,263. The repo was created 31 Aug 2025. No public affiliate program was found. cmem.ai/affiliate, /affiliates, and /partners returned 404. The "Design partners" program is TeamBrain pilots, not an affiliate. That did not change this row. Use it when you want the next Claude Code session to start from notes of the last one, and you have decided whether that compression runs on your key or on theirs. Skip it if you read the privacy page as "nothing leaves" and you will accept the default installer without reading the sign-in step. ## Basic Memory [Basic Memory](https://basicmemory.com/), from Basic Machines, is an MCP-native knowledge base. Knowledge lives as Markdown files both you and the AI can read and write. A semantic knowledge graph sits over that plain Markdown. It is not only for coding. The main limit on the paid side is the credit meter. The main limit on the docs is a stale price in the README. Agents connect through MCP, stdio or https: Claude Desktop, Claude Code, Codex, Cursor, VS Code, ChatGPT through Custom GPT actions, and anything else that speaks MCP. Obsidian can read the same files. Native packages, as of 3 Nov 2026, are a Claude Code plugin (session-start briefings, pre-compaction checkpoints, and skills) plus Hermes and OpenClaw plugins. Local storage is Markdown files plus a local index. The index can be SQLite, with Postgres as an option, and Milvus as an option for vectors. The project describes that setup as air-gapped friendly. Cloud storage is a per-tenant Tigris S3 bucket for the files and a per-tenant Neon Postgres for the index, encrypted in transit and at rest. The MCP server plus hooks come through the Claude Code plugin. The license is AGPL-3.0. The full app is open source. The [pricing page](https://basicmemory.com/pricing) was read on 3 Nov 2026. Usage plans launched 26 Oct 2026. Older seat plans are grandfathered. - Open source and self-host: free. - Cloud: $19 a month per workspace. 1,900 credits a month, 2 members, 3 API keys, 2 concurrent agents, and 7-day snapshots. A 7-day trial includes 500 credits and never bills usage. - Pro: $99 a month per workspace, billed from day one. 9,900 credits, 5 members, 10 keys. - Business: $249 a month per workspace. 24,900 credits, 25 members, unlimited keys, and audit logs. - Enterprise: custom, with on-prem. One credit is $0.01. Extra credits are the same price. Default spending caps are $38 on Cloud, $297 on Pro, and $498 on Business. SSO and a HIPAA BAA are paid add-ons. The [README](https://github.com/basicmachines-co/basic-memory) still says Cloud is $15 a month, "locked in for life," as a beta. That line is stale. Use the pricing page. Local install, from the README: ```bash uv tool install basic-memory --prerelease=allow ``` Then add the MCP server, for example `claude mcp add`, and `bm install claude-code` for the plugin. Cloud is sign in and a one-click connect. The README says `--prerelease=allow` is needed because 0.23 depends on a FastMCP 4 pre-release. The latest release, as of 3 Nov 2026, is v1.0.0 on 30 Oct 2026 IST, and it is not a pre-release. That install note may be outdated. GitHub stars: 4,462. The site says data is "never sold and never used to train AI models." The [security page](https://basicmemory.com/security) says Tigris, Neon, WorkOS, and Polar are SOC 2 and HIPAA compliant, and that Basic Memory itself does "not yet hold our own SOC 2 certification." The region is not published. No affiliate or commission program was found. The [partner page](https://basicmemory.com/partners) is invite-only, for MSPs, IT consultants, and agencies that resell or rebill workspaces. It publishes no commission figures. That did not change this row. Use it when the thing you want the agent to remember is a folder of Markdown you can open yourself. Skip the cloud plan if you need a published region, or if you are still reading the $15 line in the README. ## Mem0 [Mem0](https://mem0.ai/) calls itself "The Memory Layer for AI Agents." It is a memory API, SDK, and platform, mainly for people building apps. For coding agents it offers native plugins and a hosted MCP server. [OpenMemory](https://mem0.ai/openmemory) is Mem0's persistent MCP memory layer for coding agents. It is part of this entry, not a sixth tool. The main limit is that the coding plugins require a Mem0 Platform API key, and the per-call price after the free limits is not published. Agents, as of 3 Nov 2026: Claude Code (plugin v0.3.1, `mem0@mem0-plugins`), Cursor (lifecycle capture and six skills), Codex, Kimi Code, `DeepSeek Harness`, OpenCode, OpenClaw, Pi, and Hermes. Claude.ai connects through a remote MCP connector. Any MCP client can use the hosted server at `https://mcp.mem0.ai`, which needs a Platform API key. Skills are listed for Claude Code, Codex, Cursor, Windsurf, OpenCode, and OpenClaw. The [Claude Code doc](https://docs.mem0.ai/integrations/claude-code) says the plugin captures "evidence locally" through hooks, and that memory extraction runs through Mem0 Platform in a background worker. It exposes a local MCP tool, `search_memories`. It requires a Mem0 Platform API key. The platform stores data in a managed vector store on Mem0's servers. The self-hosted library and server (`pip install mem0ai`, or `docker compose up`) use your own stores. The OSS library and server are Apache-2.0. The platform is proprietary. Pricing changed on 26 Oct 2026. The [pricing page](https://mem0.ai/pricing) and the [pricing docs](https://docs.mem0.ai/platform/pricing), as of 3 Nov 2026: - Free: unlimited end users. For the first 30 days Mem0 pays for model calls, with limits of 10,000 adds and 1,000 retrievals a month. - Your own LLM key: those limits rise to 50,000 adds and 10,000 retrievals a month. - Past the limits: usage-based billing, with a monthly cap you set. The per-call rate is not published. - Custom: "talk to us". It names graph memory, user profiles, Dream consolidation, SSO, an SLA, and on-prem. The old Starter, Growth, and Pro prices ($19, $79, and $249) are no longer sold. Existing subscribers keep them. Accounts that signed up before 26 Oct 2026 get their 30 days until 25 Nov 2026. With your own key, Mem0 says your LLM provider charges about $3.55 to $5.34 per 1,000 adds with its recommended models. That range is the provider's charge, not a Mem0 fee. Setup, from the docs: ```bash export MEM0_API_KEY=… claude plugin marketplace add mem0ai/mem0 claude plugin install mem0@mem0-plugins --scope user --config api_key=… ``` That path needs Python 3.10 or later, and Git. The CLI is `npm i -g @mem0/cli` or `pip install mem0-cli`. The [MCP doc](https://docs.mem0.ai/platform/mem0-mcp) is the hosted server. The platform stores memories on Mem0's servers. With your own key, the text you add goes to your provider under your account. The footer says "HIPAA Ready, SOC 2 Type I, GDPR Ready." The [trust center](https://trust.mem0.ai/) is the page for those claims. OSS telemetry can be disabled with an environment variable. The region is not published. Versions, as of 3 Nov 2026: core v2.2.1 for Python, and ts-v3.3.1, both on 25 Sep 2026. Plugin releases are dated 23 Sep 2026. Last commit 31 Oct 2026. GitHub stars: 68,321, from the GitHub API on 3 Nov 2026. The site badge showed 62,590 on the same day. No official affiliate terms were found. That did not change this row. Use it when you are already building an app on a memory API and you also want a coding-agent plugin on the same account. Skip the plugin path if you will not create a platform key. The OSS server is the path that stays on your own stores. ## Supermemory [Supermemory](https://supermemory.ai/) calls itself a "Memory and context engine" and a Memory API: a memory graph, RAG, user profiles, and connectors. The default is the hosted cloud. A local mode exists, and its server source is not in the public repo. Do not call the whole product open source. The main limit is that split, plus two of the vendor's own pages that do not describe self-hosting the same way. Plugins, as of 3 Nov 2026, cover Claude Code, Muse Code, Cursor, Codex, OpenCode, OpenClaw, and Hermes. A hosted MCP at `https://mcp.supermemory.ai/mcp` uses OAuth for Claude Desktop, Cursor, Windsurf, VS Code, Claude Code, OpenCode, OpenClaw, Hermes, and ChatGPT Web. The [Claude Code doc](https://supermemory.ai/docs/integrations/claude-code) says the plugin uses SessionStart, UserPromptSubmit, PreToolUse, and Stop. It injects recalled facts and stores them in per-project containers. It needs Node 18 or later. The key is saved to `~/.supermemory-claude/credentials.json`. "Supermemory local" (`npx supermemory local`, or Docker) runs the API on port 6767 on the same machine, with local embeddings. The plugin can point at that. The [public repo](https://github.com/supermemoryai/supermemory) is MIT. The README says the plugins are open source. The [Claude plugin repo](https://github.com/supermemoryai/claude-supermemory) is separate. The [local-versus-enterprise doc](https://supermemory.ai/docs/self-hosting/local-vs-enterprise) says "Supermemory local … is free within its lite license limit of 10,000 stored documents" and "Its server source is not in the public repository." The self-hosted server you run that way is a free-tier binary, not MIT source. The FAQ lists self-hosting under Scale and Enterprise. The local binary is free up to 10,000 documents. The [pricing page](https://supermemory.ai/pricing), as of 3 Nov 2026, shows no date. It is credit-based. - Free: $0, with $5 of credits a month. - Pro: $19 a month, with $20 included, 3 seats, and auto top-up. - Max: $100 a month, with $130 included, and the Gmail connector. - Scale: $399 a month, with $600 included. SOC 2, a HIPAA BAA, a self-hosted option, and unlimited seats. - Enterprise: custom. The rate card is $5 per million SM tokens for plain-text memory and $10 per million for rich content. SuperRAG is $1 and $2 per million. Search is $5 per million queries. Operations are $100 per million. Qualifying startups and research teams get Scale free for 3 months. Subscription credits reset monthly. Top-up credits do not expire. Claude Code setup is `/plugin marketplace add supermemoryai/claude-supermemory` and then `/plugin install supermemory`. The key comes from a browser login or from `SUPERMEMORY_CC_API_KEY`. The MCP path is adding the URL. Local install, from the docs: ```bash curl -fsSL https://supermemory.ai/install | bash ``` Or `npx supermemory local`. The [privacy policy](https://supermemory.ai/privacy) is effective 29 Sep 2026. It says data may be transferred to and processed in other jurisdictions. The company is in San Francisco. The region is not published. The privacy page does not say whether customer content is used to train models. Local mode can run offline with Ollama, and with `SUPERMEMORY_DISABLE_TELEMETRY=1`. Versions, as of 3 Nov 2026: server-v0.0.9 on 12 Oct 2026. Last commit 3 Nov 2026. GitHub stars: 31,513 on the main repo, and 2,798 on the Claude plugin repo. Supermemory's own pages claim it "ranked top on the independent SWE-ContextBench benchmark" in February 2026, that it handles "100B+ tokens a month," and that median recall is "187ms." Those figures are the vendor's claims. A public partner page exists. The link in this section is the product site. That did not change this row. Use it when you want a hosted memory API and you have read the credit card, or when the 10,000-document local binary is enough and you do not need the server source. Skip it if "open source" has to include the process you run on your own machine. ## Memory tools for AI coding agents that were left off - [Letta](https://letta.com/). Apache-2.0. Left off because the product is itself a coding agent, not a memory add-on for agents you already run. - [Zep](https://getzep.com/), including Graphiti. Apache-2.0. Left off because it is a temporal knowledge-graph library for app builders who need a graph database. No coding-agent plugin was part of this check. - [Cognee](https://github.com/topoteretes/cognee). Apache-2.0. It is a general memory and knowledge-graph platform for agents and apps, and it ships Claude Code and Codex plugins. It is the strongest candidate to swap in. Left off because the focus is broad, rather than coding sessions. Re-check it at the next refresh. - [agentmash](https://agentmash.dev/) is not a memory tool. It belongs in the [multi-agent workspaces post](https://www.dharmikjagodana.com/blogs/best-multi-agent-coding-workspaces-for-solo-builders.md). - OpenMemory is Mem0's product. It is covered in the Mem0 section. Session logs are the thing these tools store. I wrote about what not to leave exposed in [I Exposed My AI Agent to the Internet: An OpenClaw Security Guide](https://www.dharmikjagodana.com/blogs/how-to-secure-openclaw-on-cloud.md). That post is about OpenClaw. It is not a test of the five tools here. ## How to choose Pick by where the session is allowed to go, and by whether you want notes, files, or an API. - You want a local log of many coding agents, and you will opt out of the cloud during setup. Beacon. Read the research-use section before you leave Beacon Cloud selected. - You want the next session to start from notes, and you will decide who runs the compression. claude-mem. Do not treat the privacy page and the default installer as the same promise. - You want a folder of Markdown you can open, with or without their cloud. Basic Memory. Ignore the $15 line in the README. - You are building an app on a memory API and you will create a platform key for the coding plugin. Mem0. The old $19, $79, and $249 plans are not the current offer. - You want hosted credits, or the free local binary is enough and you do not need its source. Supermemory. The local binary is free up to 10,000 documents. The FAQ lists self-hosting under Scale and Enterprise. If the agent also runs code, where that code runs is a separate choice. That list is [5 Best AI Agent Sandboxes for Solo Builders](https://www.dharmikjagodana.com/blogs/best-ai-agent-sandboxes-for-solo-builders.md). If two of those fit, use the one whose default path matches where you are willing to send a session. A local default that the installer then overrides is not the same as a local default you keep. ## FAQ ### Does Beacon send a coding session to the cloud by default? Local use, once you are on the local destination, sends nothing to them. The interactive setup preselects Beacon Cloud, and Local is an opt-out. On Beacon Cloud, research use is on unless you opt out. That opt-out takes effect within 30 days, and models already trained are not retrained. The standard cloud region is not published. Enterprise can name its own bucket and region. ### Does the claude-mem privacy page match the default install? Not on the point that matters. The privacy page says nothing leaves the device unless you turn on cloud sync, and that memory content is not used to train models. The README says the installer asks you to sign in by default, which turns on the hosted observer for a 30-day trial. Session content may go there even with cloud sync off. The privacy page does not address that. ### Is Basic Memory Cloud still $15 a month? Not on the pricing page read on 3 Nov 2026. Cloud is $19 a month per workspace. The GitHub README still says $15 a month, "locked in for life," as a beta. Use the pricing page. Usage plans launched 26 Oct 2026. Older seat plans are grandfathered. ### Does Mem0 still sell the $19, $79, and $249 plans? No. Those Starter, Growth, and Pro prices are no longer sold, as of the 26 Oct 2026 change. Existing subscribers keep them. The current page is a free tier, then usage billing with a cap you set. The per-call rate is not published. Accounts that signed up before 26 Oct 2026 get their 30 days until 25 Nov 2026. ### Can Supermemory run without their cloud? Yes, as "Supermemory local," free up to 10,000 stored documents, on port 6767 on the same machine. The docs say that server's source is not in the public repository. The pricing FAQ also says self-hosted deployments are a Scale and Enterprise option. Those are two different claims. The public repo and the plugins are MIT. That license does not cover the local server binary. ### Which memory tool works fully offline? Basic Memory's local setup is Markdown files plus a local index. The project describes that setup as air-gapped friendly. Open source and self-host are free. Beacon's local destination sends nothing off the machine. The interactive setup preselects Beacon Cloud. Local is an opt-out. Supermemory local can run offline with Ollama, free up to 10,000 stored documents. The server source is not in the public repository. Mem0's self-hosted library and server use your own stores. The coding plugins require a Mem0 Platform API key. claude-mem stores notes in local SQLite and Chroma. The default installer asks you to sign in, which turns on the hosted observer. New posts show up on X. [Follow new posts on X](https://x.com/dharmikjagodana) [All posts](https://www.dharmikjagodana.com/blogs.md) ## Related - [5 Best Tools for Running Multiple AI Coding Agents in Parallel (2026)](https://www.dharmikjagodana.com/blogs/best-multi-agent-coding-workspaces-for-solo-builders.md): Tools for multiple AI coding agents, compared on price, MCP, and which CLIs they run. Five managers for a solo builder, from pages checked Oct 26, 2026. - [5 Best Computer-Use Tools for AI Agents (and How to Run Them Safely) (2026)](https://www.dharmikjagodana.com/blogs/best-computer-use-tools-for-ai-agents.md): Computer use tools for AI agents, compared for builders who let a model use a screen. Five vendor options and the shared safety rules. Last checked 2 Nov 2026. - [5 Best Tools That Give Your Coding Agent Eyes on Your App (2026)](https://www.dharmikjagodana.com/blogs/best-app-context-tools-for-solo-builders.md): App context tools for solo builders, compared by what the agent sees and whether that data stays put. Jam, Faylee, ilse, Reticle, and Chrome DevTools MCP.