Your agent keeps relearning the same small facts.
Stable preferences, project facts and workflow rules matter again next session, but replaying the old conversation is wasteful.
Free · local-first · governed memory for AI agents
What should your agent remember? And what should it never treat as truth?
FactLane adds a small governed fact lane beside the memory you already use. Keep useful preferences, project facts and workflow rules across sessions with scope, provenance, freshness and Candidate → Current verification.
- Apache-2.0 open source
- v0.1.3 production-qualified local profile
- Five tools focused MCP surface
Prefer manual setup? Open the Quick Start →
FactLane · governed factsLOCAL
YouRemember that I prefer concise answers in Arabic, but keep technical terms in English.
After trusted verification · PreferenceCURRENT
Concise Arabic explanations + English technical termsscope · global usersource · user instruction
Next week
Your assistant can reuse the approved preference without rebuilding the context from scratch.
Quick fit check
Is this a FactLane problem?
FactLane is useful when a small fact should survive a session, but you still need to know where it came from, how fresh it is and whether it is trusted as Current.
Remembered facts can become stale or misleading.
You want provenance, freshness and an explicit Candidate → Current lifecycle instead of treating every saved fact as equally trustworthy.
You want memory without giving memory authority.
Current instructions, live project state and verified live sources must still outrank anything the agent remembers.
What is FactLane?
A governed fact layer for AI agents that need memory across sessions.
FactLane is a free, open-source, local-first MCP memory layer for compatible AI-agent hosts. It keeps small reusable facts with scope, provenance and freshness; a contribution can be a Candidate without becoming verified Current state, and current instructions, live project state or verified live sources still win.
What problem does it solve?
An agent can reuse a useful preference, project fact or workflow rule without replaying the old conversation — and without treating stale memory as current truth.
How is it different from built-in memory?
Broad or native memory is good for rich recall and continuity. FactLane is the narrower governed lane for facts that need exact scope, provenance, freshness and Candidate → Current verification.
What does it not replace?
It does not replace your assistant, broad memory, a knowledge base or execution authority. It sits beside them and governs a small class of reusable facts.
Start with your agent, not a terminal
Give FactLane to the agent you already use.
Ask your agent to read the project, compare it with your current memory stack and explain whether the governed-fact model solves a real problem for you. Installation comes after understanding fit.
Evaluate fit first
Ask what FactLane adds beside the memory you already have, which facts belong in it, and which information should stay in broader memory or knowledge systems.
Learn the trust model
Understand Candidate, Current, freshness, scope and why remembered information remains below current instructions and live state.
If the agent has access, install read-only
Use the exact v0.1.3 release, the supported local profile and a compatible MCP host. Verify all five tools and a suitable memory_status call before enabling writes.
Enable contribution only after the boundary is clear
If Candidate writes are useful, grant only the delegated-candidate profile. An ordinary agent does not grant itself verifier authority.
What belongs in the governed lane?
Keep the facts that should survive a session.
FactLane is strongest when a small fact will matter later and you want to know where it came from, how fresh it is, and whether it is actually Current.
Project state
Carry forward facts that change how work should continue.
“The release is v0.1.3. Production changes still require owner approval.”
Current retrieval can reuse the verified fact without replaying the project history.
Stable preferences
Remember durable working preferences without making them absolute.
“Keep answers concise unless I explicitly ask for a deep explanation.”
The preference can survive the session while a new current instruction still wins.
Workflow boundaries
Preserve rules that agents need to see again.
“Production changes need approval. Local tests can run automatically.”
A bounded fact can be reused across compatible hosts without turning the memory store into the authority that grants execution rights.
Complement, do not replace
Keep your broad memory. Add FactLane as the governed fact lane.
Integrated memory is not one problem with one correct architecture. Different layers are good at different jobs.
Rich recall and continuity
Useful for preferences, summaries, long-lived context and the broad personal or agent memory experience.
Connected knowledge at scale
Useful for relationships, documents, concepts and large bodies of organized knowledge.
A governed fact lane
Useful when a small fact needs exact scope, provenance, freshness, Candidate → Current verification and explicit authority boundaries.
Codex and Hermes are tested stdio hosts. The intended pattern is coexistence: FactLane can sit beside the other memory layers in those workflows rather than trying to replace them.
Useful memory without hidden authority
Memory can help without becoming permission.
FactLane separates memory eligibility from execution authority. This is the center of the design, not a warning added afterward.
Current reality wins.
A current user instruction, live repository/product state or verified live source outranks memory when they conflict.
Contributors do not verify themselves.
A normal delegated agent can contribute a Candidate when allowed; trusted promotion is a separate operation with revision and identity checks.
Write authority belongs to the launcher boundary.
Seeing memory_store or memory_update does not itself grant permission to use them with privileged semantics.
CandidateWorth remembering, but not trusted as current yet.
Trusted verificationA separate trusted step checks identity, revision and lifecycle.
CurrentVerified, fresh enough and eligible for current-state retrieval.
Quality you can inspect
Free and open source. Production-qualified where we claim support.
FactLane v0.1.3 is an official production release for a deliberately bounded local profile. The quality claim is tied to explicit contracts, fail-closed behavior and documented qualification — not adjectives.
Small public surface
Exactly five MCP tools keep the contract focused: search, read, contribute, governed update and status.
Fail-closed defaults
Read-only is the default. Unsupported runtime/provider conditions and unauthorized transitions fail closed instead of silently degrading into broader authority.
Production qualification
The documented profile was exercised for installation, backup/restore, bounded concurrency, crash/restart behavior, configured host startup, production-derived retrieval and SQLite capacity failure.
Inspectable and local
Apache-2.0 source can be inspected, run and adapted. Local SQLite/SQLite-vec storage and supported local Ollama embeddings mean the qualified profile does not require a hosted memory service or external embedding API.
The boundary is explicit: stdio MCP, Python 3.11+, linked SQLite 3.42.0+, supported local Ollama embeddings and the documented local POSIX storage/recovery contract.
A governed lane for reusable facts