Agents that remember.
Long-term memory for Claude Code, Codex and any agent. Plain files you own.
How memfor remembers: write, index, recall.
-
Write.
Every lesson your agent verifies becomes a file. So the next session doesn't start from zero.
--- id: learn-ci-cache-key category: learnings tags: [ci, cache] source: agent --- A stale dependency cache passed CI. Hash the lockfile into the cache key.An example learning, in the format the toolkit writes. -
Index.
The toolkit files it into a deterministic index. No LLM, no rewrite. Same files, same ranking.
- learn-ci-cache-key learnings
- learn-retry-idempotent-only learnings
- skill-release-checklist skills
- project-payments-api projects
An example index. Only the toolkit writes it. -
Recall.
Ask, and the right five come back, inside a token budget.
How do I share one knowledge base with my team?
- 1 docs/GUIDE.md › Syncing the knowledge base over git (team mode)
- 2 docs/loop.md › Three Phases
- 3 docs/GUIDE.md › A typical session
- 4 docs/GUIDE.md › How the memory is built (and how to use it)
- 5 docs/GUIDE.md › The cadence
Recorded output from the toolkit at a pinned commit.
One memory engine. Four ways to use it.
- mem for AI Memory for any agent.
- mem for Code Long-term memory for Claude Code and Codex.
- mem for Teams One shared memory for your team. Early access · not built
- mem for Learning Memory for how you learn. Exploring · no product yet
Yours, readable, auditable. Memory you can open in an editor.
-
Plain files.
Markdown and JSON in a folder you choose. Back it up, diff it, delete it. -
No LLM in the toolkit.
Identical rankings across runs for a fixed knowledge base and date. -
Git is the audit trail.
Every change is a diff you can read, and you choose the remote.
Recall@5 on LongMemEval. 470 questions, scored at a pinned commit, with the method and the limits published.
Start in your terminal. Choose where its memory lives.
Clone it, run your agent, and answer a few questions. Your knowledge folder stays on your machine.
Paste setup into your agent
Set up agentware for me. Clone https://github.com/r5rana/agentware.git into ./agentware and check out commit 3c25cddd28b5e4d1c320c1e34a98eacaecf6a199. Read ./agentware/README.md and follow its onboarding. Stop and ask me before you install anything, before you run a command that writes outside ./agentware, and before you choose where my knowledge base lives. Stay inside your sandbox. If onboarding offers the PLAN_AW and WORK_AW shell aliases, or offers to start the unattended loop, ask me first: both start you with approvals turned off. If a step fails, tell me instead of retrying with broader permissions.
This text is shown in full here. Nothing is hidden in the clipboard.
git clone https://github.com/r5rana/agentware.git
cd agentware
git checkout 3c25cdd
claude "hi"git clone https://github.com/r5rana/agentware.git
cd agentware
git checkout 3c25cdd
codex "hi"This installs agentware, the open-source engine behind memfor. No tagged release yet. These commands pin commit 3c25cdd.
Apache-2.0. No memfor account. Needs macOS or Linux (Windows via WSL), git, Python 3, bash, jq, and Claude Code or Codex.
What your security team will ask. Answered before they do.
Where does memory live?
In a folder you choose, as Markdown and JSON. Back it up, diff it, delete it.
Does anything phone home?
agentware sends no telemetry and has no account system. Your agent runtime talks to its own model provider.
What can write to memory?
Only the toolkit writes the index, and learnings enter memory through it. The files are yours: anything with access to the folder can edit them, and git shows the diff. Changes to agentware itself go through an adversarial review gate.
Can an MCP client change my knowledge base?
No. The MCP server is read-only.
How do I verify what I install?
Today you pin a commit. The first signed release tag is planned, and verification steps will ship with it.
Is team sync safe?
It's experimental. We're hardening the trust boundary for synced content, so sync only with collaborators you trust.
Does the loop run my agent unattended?
Yes. The loop and the optional PLAN_AW and WORK_AW shell aliases launch Claude Code with --dangerously-skip-permissions (Codex: --dangerously-bypass-approvals-and-sandbox). Run it in a container, a VM or a dedicated account. Onboarding asks before it installs the aliases.
