What each tool is
ai-memory, by Fabio Akita, is an MIT-licensed long-term memory server for coding agents. Lifecycle hooks and MCP registration record what happens in a session, sanitize it and consolidate it into readable markdown pages in a git-backed, per-project wiki. The next session, in any supported agent, gets a short brief and can search earlier work. It ships as one self-hosted Rust binary; the latest release is v3.1.2 (October 20, 2026).
Beacon, by Asymptote Labs, is an MIT-licensed layer that sits across your coding agents. Hooks, plugins and adapters capture full session history from Claude Code, Cursor, Codex, OpenCode, Cline and more than 20 other harnesses, normalize it into one OpenTelemetry-based event model, and turn reviewed workflows, corrections and debugging patterns into skills and memory that future agents retrieve through MCP or Agent Skills. The latest release is v1.3.45 (October 20, 2026).
How memory is captured and stored
ai-memory treats a git-backed folder of ordinary .md files as the source of truth. You can grep it, open it in Obsidian or edit it by hand; the search index is derived from those files and can be rebuilt. Retrieval fuses full-text, entity and link search, with optional local vector search. The default path makes no LLM calls; LLM consolidation and an opt-in "dream" pass that merges cold notes are off by default.
Beacon stores a full trace of each session: prompts, responses, tools, commands, edits, approvals, MCP activity and tokens. It writes endpoint telemetry to local JSONL by default and can export to Elastic, S3 or any OpenTelemetry collector. Memory is a layer on top of that history: Beacon surfaces findings (security risks, reusable skills and behavior patterns) and turns the ones you keep into skills and MCP-accessible memories.
Moving work between agents
ai-memory's headline feature is the cross-agent handoff. Quit Claude Code mid-task, open Codex in the same directory, and the next agent receives a typed handoff describing where you stopped, what failed and what is still open; a handoff has an owner and is claimed once. Beacon's model is broader and less explicit: because it normalizes history from every harness, knowledge learned in one tool can be reused by another, and agents can search past sessions over MCP, but it does not describe a claim-once handoff protocol.
Observability and security findings
This is where the two differ most. Beacon is also an agent observability and security product: it gives you a local terminal and web dashboard for traces, event timelines and token usage, and Beacon Cloud flags risky actions such as a secret pasted into a prompt, a destructive command, an agent bypassing its permissions or a call to an unknown MCP server. Enterprise adds the option to block or require approval for risky actions and fleet rollout through Jamf, Intune or Fleet. ai-memory is a memory and handoff tool; it does not position itself as a trace viewer or a security scanner.
Supported agents and platforms
ai-memory runs on Linux and macOS, on Windows via WSL2, and experimentally on native Windows, with a Claude Code plugin, AUR and RPM packages, a macOS app and a Docker image. Its support matrix lists first-party integrations for Claude Code, Codex, Cursor, Gemini CLI, OpenCode and other agents, some of them MCP-only. Beacon installs through Homebrew on macOS, .deb or .rpm packages on Linux and an MSI on Windows, and its README lists more than 20 supported harnesses plus browser agents, cloud agents, CI workflows and agent SDKs.
Privacy and data handling
ai-memory is self-hosted only. A default install listens on 127.0.0.1, and its security page says it has no analytics, crash reporting or usage tracking. Beacon is local-first but has a cloud default in its interactive setup: the wizard preselects Beacon Cloud, with an explicit Local option. Signing in forwards nothing; confirming the Cloud option installs Beacon and starts forwarding session data to Asymptote Labs. Choose Local to keep everything on your machine. Noninteractive installs are account-free.
Pricing
ai-memory is free and open source, with no paid tier; team features such as multi-user auth, per-person attribution and an audit log are included. You pay only for optional model or embedding providers you connect. Beacon's open-source stack is free, and Beacon Cloud has a free tier with 50 findings. Cloud Pro costs $20 per month, or $15 per month billed annually, and adds unlimited findings, full MCP access, agent memories and skills pulled from sessions. Enterprise is custom-priced.
Who should pick which
- Pick ai-memory if you want memory you own as plain markdown in git, explicit handoffs between agents, team sharing on your own server and no cloud component at all.
- Pick Beacon if you also want to see what your agents did across many harnesses, get security and habit findings, and are comfortable with a managed cloud for the memory features (or will self-host the open-source trace stack).
- Use both if you want Beacon's traces and findings alongside ai-memory's wiki and handoffs; nothing in either project's documentation prevents running them side by side, but we have not tested that combination.
The Bottom Line
ai-memory is the simpler, fully self-hosted memory and handoff tool. Beacon is a wider trace, findings and memory platform with a paid cloud tier.