The remnant that survives.

mem + remnant. What's left after context windows die, conversations end, and you come back three weeks later. Memory is what happened; institutional knowledge is what is still true. A local SQLite decision ledger that knows what you decided, when, why, what replaced it, and whether it still holds.

What your agent gets
recall
Semantic search over past decisions, session logs, and framework fixes. Includes cross-project colony results marked [colony].
log
Record decisions, framework fixes, and session summaries. Framework fixes and rejections auto-promote to your colony.
session_context
Compiled context with narrative briefing: last session, open TODOs, stale decisions, colony fixes. LLM-composed or template.
session_log
Log a progress summary without closing the session. Default when you say "done".
session_close
Close a session with a summary of what shipped, what was decided, and what's next.
synthesise
Ask questions that span multiple records, with citations back to source.
context_for_file
Records relevant to the file being edited.
project_brief
Dynamic 500-token summary of where the project stands right now.
status
Project health: stale records, unresolved contradictions, ledger stats.
check_copy
Validate text against your copy specs. Catches banned phrases.
check_design
Validate against design system specs. Catches banned components.
promote
Manually promote any project record to the colony for cross-project access.
analytics
Decision velocity, knowledge areas, coverage gaps. Sparkline dashboard or JSON.
team_status
Team health: active builders, record counts per builder, unresolved contradictions, last import timestamp.
brief --onboarding
Structured onboarding package: key decisions, architecture patterns, known gotchas, team conventions. 8K token budget.
stigmergy
Reactive cross-builder awareness. When a teammate logs a fix for a file you're working on, it surfaces immediately, not next session.
churn detection
Spots topics that keep getting revisited. Decisions superseded 3+ times raise a churn alert.

Full list on npm.

How it works

Typed records (decisions, framework fixes, session logs) live in an append-only SQLite ledger with vector embeddings for semantic search. Decisions are never edited: a new record supersedes the old one, and the chain stays visible. When the codebase changes under a decision, memnant scores whether the change actually matters: semantic confidence, not a binary flag. Everything above runs locally; only synthesis calls out, to Anthropic, OpenAI, or a self-hosted model via Ollama.

Who it's for

You work in sessions. Days or weeks apart. Across projects that share patterns. memnant is for builders who want what they learned yesterday to be there tomorrow, and still true.

$ memnant

how to install · works offline · mit license