memfor
GitHub
Display
Theme
Motion
memfor
Theme
Motion 
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.

Send setup steps to my computer Join the Teams waitlist

mem for AI

Memory for any agent.

The engine behind memfor. Deterministic recall over files you own.

Send setup steps to my computer Read the quickstart

01

Write once, recall anywhere. Any agent can ask.

agentware gives AI agents a long-term memory and a self-learning work loop. Knowledge is stored as plain Markdown and JSON in a directory the user owns, and it runs on top of an agent runtime you already use, Claude Code or OpenAI Codex, instead of replacing it.

agentware recall

Where is my knowledge base stored, and can I move it to another machine?

  1. 1 docs/loop.md › Three Phases
  2. 2 README.md › Reliable Agentic Loops with persistent memory
  3. 3 docs/GUIDE.md › A typical session
  4. 4 docs/GUIDE.md › Syncing the knowledge base over git (team mode)
  5. 5 docs/GUIDE.md › You own your data
Recorded output from the toolkit at a pinned commit.

02

Deterministic by design. Same files, same ranking.

Recall is BM25 ranking with a token budget, and there is no LLM in the toolkit. For a fixed knowledge base and date, the ranking is identical across runs, so you can benchmark it and diff it.

Run 1

How do I share one knowledge base with my team?

  1. 1 docs/GUIDE.md › Syncing the knowledge base over git (team mode)
  2. 2 docs/loop.md › Three Phases
  3. 3 docs/GUIDE.md › A typical session
Run 2

How do I share one knowledge base with my team?

  1. 1 docs/GUIDE.md › Syncing the knowledge base over git (team mode)
  2. 2 docs/loop.md › Three Phases
  3. 3 docs/GUIDE.md › A typical session

The recorder ran this query twice at a pinned commit and fails unless both runs match.

03

Works over MCP and the CLI. No account, no server to host.

Ask from the command line, or serve the knowledge base to any MCP client. The MCP server is read-only: only the toolkit writes the index.

zsh — agentware
scripts/agentware recall "How do I share one knowledge base with my team?" --format json1  docs/GUIDE.md › Syncing the knowledge base over git (team mode)2  docs/loop.md › Three Phases3  docs/GUIDE.md › A typical sessionscripts/agentware mcp serve# serves your knowledge base to MCP clients over stdio; it can't write
Commands from the CLI; recall results recorded at a pinned commit, abbreviated.

Everything in the engine. Each with its status.

  • CLI recall

    Available

    Returns ranked entries within a token budget, as text or JSON.

    scripts/agentware recall "<query>" --format json

  • Read-only MCP server

    Available

    Serves your knowledge base over stdio to MCP clients. It can't write.

    scripts/agentware mcp serve

  • Claude Code and Codex

    Available

    The plan, execute, verify, learn loop runs on the agent you already use.

  • Semantic recall (Mode B)

    Opt-in

    BM25 plus a pinned, on-device embedding model.

  • Local models (LM Studio)

    Experimental

    Run the execute phase on a local model while planning stays on your cloud runtime.

  • Plain files, one writer.

    Available

    Knowledge is plain Markdown and JSON in a directory you own. A deterministic toolkit is the only writer of the index. Licence: Apache-2.0.

Specifications. Measured, with the test conditions.

Characteristics
Parameter Value Test conditions Evidence
Recall@5 0.9140 measured LongMemEval-S (cleaned), 470 answerable questions, session granularity, BM25, top 5, default mode Ledger row 86 · 42d58a0
nDCG@5 0.8831 measured Same run Same row
MRR 0.9104 measured Same run Same row
Determinism Identical rankings and metrics Fixed knowledge base and --as-of date; latency excluded test_recall_json_is_byte_identical_across_runs
Recall defaults top-k 5, token budget 1,500 n/a recall --help at 3c25cdd
Third-party Python packages 0 (default mode) Standard library only. Still needs bash, jq, git and Python 3. Mode B adds one pinned package. test_cli_imports_are_stdlib_only
LLM calls in the memory path (init, learn, index, recall, query, mcp) 0 Your agent is an LLM; the memory path isn't. The review gate and the opt-in checkup run your agent's CLI. README ("zero LLM in the memory path") and source
Vetted skills shipped 14 catalog/skills.json at 3c25cdd Catalog
Latency See the benchmark ledger Hardware not yet recorded /benchmarks#latency

Give your agent a memory.

Send setup steps to my computer View source on GitHub 
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.

Operating requirements
Claude Code or OpenAI Codex · macOS or Linux (Windows via WSL) · git, bash, jq, Python 3.
Revision history
Changelog
Licence
Apache-2.0 with NOTICE
Esc