mem · remnant · ant

It just remembers.

You work across projects, across tools, across weeks. memnant carries what you decided from one session to the next, from one project to every project.

$ memnant
how to install · works offline · mit license
scroll

Context windows die. Conversations end. Three weeks later you're re-explaining the same decisions to the same agent. memnant is the remnant that survives.

01—05 The core loop
01
It
remembers.

The agent logs silently as you work. Decisions, framework fixes, rejections, product calls. Every record gets a vector embedding for semantic search. You don't do anything — it just remembers.

02
It
compiles.

Next session starts with what matters. Last session's summary, open TODOs, relevant decisions, staleness warnings. Not a raw dump — the relevant subset, with what's changed since you were last here.

03
It
knows.

It knows when knowledge goes stale. When a file changes, memnant scores whether the change actually affects related decisions. Semantic confidence, not a binary flag. A renamed variable doesn't invalidate an architecture decision.

04
It
travels.

Knowledge crosses projects automatically. Framework fixes and rejected approaches promote themselves to a machine-local colony. The gotcha you solved in one codebase is already waiting in the next.

05
It
shares.

Multiple builders, one shared understanding. Knowledge logged by one builder is available to all. Contradictions detected automatically. New team members get an onboarding brief — key decisions, patterns, gotchas — in minutes, not days.

In practice
memnant — session start
$

More than memory.

Remembering is the start. What makes knowledge useful is how it connects, what fades, what stays, and what gets enforced.

Connection graph
Records link themselves by semantic similarity. Supersession chains track when new decisions replace old ones. Contradictions are flagged automatically.
Relevance decay
Old knowledge fades. Frequently accessed records stay prominent. The ledger self-organises around what's actually useful.
Synthesis
Ask questions that span multiple records. "How did our auth approach evolve?" returns a composed answer with citations back to source.
Governance
Spec enforcement in pre-commit hooks. When overrides pile up, memnant suggests updating the spec — the rule might be wrong, not the code.
Colony
A machine-local ledger at ~/.memnant/colony.db. Framework fixes and rejected approaches auto-promote across all your projects. Deduplicated by embedding similarity.
Narrative briefings
Session context rendered as a story, not a database dump. Template-driven offline. LLM-composed when an API key is available. Graceful fallback.
Team layer
Multiple builders share one ledger. Decisions sync through git. Contradictions detected on import. Cross-builder recall, team patterns, and an onboarding brief that gets new members up to speed in minutes.
Pheromone trails
Records accessed together get boosted together. The more you look at two decisions side by side, the stronger the trail. Stale trails decay after 90 days.
Stigmergy
When a teammate logs a fix for a file you're working on, it surfaces immediately — not next session. Contradictions flagged in real time.
Churn detection
Spots decisions that keep getting superseded. When a topic has been revised 3+ times, memnant flags the underlying tension instead of accepting another revision.
Self-hosted models
Point synthesis at Ollama, LM Studio, or vLLM. One config field. Core features work fully offline — no API key needed for search, staleness, or context.
Cross-project synthesis
Ask questions that span your colony. "What auth patterns have I used?" searches across every project on your machine, with citations.
Under the hood
Storage
Single SQLite file. .memnant/ledger.db — copy it to another machine and everything comes with you.
Search
Local vector embeddings via all-MiniLM-L6-v2. Semantic search on CPU. No API calls. Works on a plane.
Integration
MCP server over stdio. Plugs into Claude Code, Cursor, or any MCP-compatible agent.
CLI
memnant — one command. Auto-detects: first run → setup, after that → session start.
Config
memnant.yaml at your project root. Version-controlled. No dashboard. No account. No login.
Export
Markdown or JSON. Every record, every decision, every session log. Your history is never locked in.

Start building.

One command. Full context. The agent already knows the rest.

$ memnant

how to install · no dependencies · works offline