Cogni-Code — Persistent memory for AI agents

Give your agent a memory that outlasts the session.

Every new session starts from zero — forgetting your preferences, decisions, and the bug you spent hours on yesterday. Cogni-Code fixes that with a persistent, inspectable knowledge graph that learns how you work across sessions.

$ npm install -g cogni-code

A fourth option for agent memory.

Most agent memory is built-in and opaque, locked in a vector database, or hand-written by you. Cogni-Code is something else: plain files on disk, maintained automatically, capturing behavior — not just facts.

	Cogni-Code	Built-in memory	Vector-DB memory	Hand-written files
Where it lives	Your filesystem	Vendor cloud	Database / SaaS	Your repo
Can you read it?	Yes — cat, grep, diff	Limited UI	Partial	Yes
Self-hosted	Yes	No	Sometimes	Yes
Maintained by	Background pipeline	Vendor	Agent or app	You, manually
Captures	Behavior + facts	Mostly facts	Mostly facts	Whatever you write
Decays when stale	Yes	No	Usually no	No
Multi-agent	Yes — Claude, Codex, OpenCode, pi	No	Sometimes	Manual
Writes its own tools	Yes — Skillforge	No	No	No
Git-backed history	Yes	No	Rarely	If you commit it

Capture, process, inject, evolve.

You just talk to your agent. Session hooks capture state into a local buffer. A separate background pipeline distills it into memory, and your next session starts already knowing who you are and where you left off.

Phase 01

Capture

Session hooks watch conversations and tool traces. Nothing is sent anywhere — hooks write to a local buffer.

Phase 02

Process

The scribe extracts deltas, the auditor finds stale nodes, the librarian applies judgment-heavy updates, the dreamer makes associations.

Phase 03

Inject

The next session starts layered: who you are, how this project works, what happened recently, what's most relevant. A few thousand tokens.

Phase 04

Evolve

Memory decays when unused, dreams surface unexpected connections, and Skillforge promotes frequently-recalled memories into commands.

Watch a session flow through the pipeline.

It runs in the background, so your agent never manages memory mid-conversation. Each stage lights up as memory passes through — scribe to dreamer is the active chain.

scribecore

Reads the session buffer and extracts structured deltas — evolving opinions, frustrations, and contradictions, not just hard facts.

delta · session 2026-07-15
+ prefers blue-green deploys for prod   0.82
+ frustrated by silent deploy waits      0.61
~ decisions/deploys: canary-everywhere  stale

auditorcore

Mechanical triage — scans the whole graph for stale, contradictory, and noisy nodes.

triage · 412 nodes
✗ 6 stale      confidence < 0.3, 34d untouched
✗ 2 contradicts decisions/auth-strategy
~ 11 bloat     candidates flagged for prune

librariancore

Applies judgment-heavy updates with a prune-over-preserve philosophy, then regenerates the context files.

applied
3 upserts · 2 archived · 1 merged
→ regenerated MAP.md      174 entries
→ regenerated WORKING.md
commit a1f3c9e

dreamercore

Recombines distant nodes into speculative fragments for later validation.

fragment · emergence  0.57
“single-custody flows” — reminder
delivery, cron restore, and the token
cache all broke on shared custody.
→ linked 3 nodes

observeraux

Turns conversation patterns into structured observations, session logs, and node upserts.

observations
+ mind/observations.jsonl
  ships by vertical slice, not feature flags
+ sessions/cogni-code.jsonl
  shipped, decided, blocked, next

skillforgeaux

Promotes frequently-recalled nodes into executable slash-command skills.

recall heat
“ssh provisioning”  × 8 sessions
→ generated /provision-ssh
→ installed ~/.claude/commands

bootstrapaux

Auto-generates project docs from mental-model data.

lens · cogni-code
+ tech stack, conventions, active threads
→ wrote ./CLAUDE.md  4.2 KB
→ wrote ./AGENT.md   codex

Give your agent a memory that outlasts the session.

Works with the agent you already use — on both ends.

Switch your coding agent

Switch your memory worker