Loading
Loading
Team memory for coding agents
Human-approved skills, visible sources, and outcome receipts, loaded before your coding agents guess.
Cursor, Codex, Gemini, and Claude Code each get the matching installer, then the agent verifies assist, save, and outcome reporting before work starts.
solo
skills, decisions, installs
team
shared skills with authors
proof
reuse receipts and outcomes
Founder and team background includes work at:
For individuals
Personal Cognition keeps preferences, decisions, fixes, and project scope available when a new chat starts.
Generate scoped keyFor organizations
Teams approve reusable workflows, keep author attribution, and review outcome receipts before trust compounds.
Start team pilotfirst run
Cognition should never be opaque. A good run shows what triggered it, what it loaded, when it had no match, and whether the reused skill helped.
setup contract
User, team, host, and project scope are explicit.
Codex, Cursor, Gemini, Claude Code, or any MCP agent.
Confirm assist, ask, browse, forget, save, and outcome reporting are loaded.
Every reused skill ends as helped, refreshed, or incomplete.
Install
Adds MCP config and local instructions.
Assist
Returns the matching skill or a clear no-match.
Approve
Trigger, steps, checks, owner, failure modes.
Reuse
Future agents get the workflow before guessing.
run contract
Call ledger
Trigger, loaded sources, action taken, result, and timestamp.
Fast no-match
If context is thin, the agent asks for missing input instead of guessing.
Project context
Preferences and anchor docs stay scoped to the active project.
Excluded by default
Secrets, raw source, private keys, full transcripts, and drafts.
proof users can inspect
Users should not have to infer why Cognition activated. Every assist call needs to show the trigger, source boundary, answer, and outcome, or return a fast no-match when context is too thin.
Visible triggers
Users should know why Cognition activated, not infer it from chat history.
Low-signal fallback
A fast no-match beats a slow, ungrounded answer.
Persistent preferences
Style, repo rules, and project goals should survive new chats and context reshuffles.
Cross-chat continuity
The win users feel: less repeated setup, fewer repeated corrections.
sample run receipt
helped: truetrigger
Task intent, active files, agent host, and project scope.
sources
Approved skills, author, freshness, and source boundary.
answer
Use the skill, ask for missing context, or return no-match.
outcome
Marked helped, refreshed, or incomplete after the run.
In the local MCP runtime, this maps to `call_ledger`, `low_context`, saved `activity_events`, and sampled outcome prompts instead of another dashboard chore.
how it compounds
SKILL LOOKUP
Ask first
Step 01At task start, the agent checks what your team has already figured out.
EVIDENCE LOG
Capture work
Step 02Commands, file edits, stuck points, and outcomes become evidence for a reusable workflow.
DRAFT SKILL
Save skills
Step 03Cognition drafts the SKILL.md and waits for a human yes before sharing it.
AGENT CONTEXT
Retrieve later
Step 04Bob hits the same wall later. His agent loads Alice's fix before guessing.
Run a one-week pilot: install Cognition, capture real skills, and prove one teammate unblock.
See the builder pilotvs. the alternatives
Mem0, Zep, and Letta pushed memory forward. Cognition is narrower: coding-agent procedure that a human approved, an author taught, and a later run can verify.
Mem0
Broad user and app memory for personalized agents.
Zep
Temporal graph memory for enterprise agent context.
Letta
Stateful agents and context management runtime.
Cognition
Governed coding-agent skills: approval, attribution, freshness, receipts.
| Feature | Context files | RAG / embeddings | ✓Cognition |
|---|---|---|---|
| Persists across sessions | ✓ | ✓ | ✓ |
| Human approval before team sharing | — | — | ✓ |
| Author attribution on every skill | — | — | ✓ |
| Outcome receipts after reuse | — | ~ | ✓ |
| Readable outcome receipt fields | — | — | ✓ |
| Fast no-match when context is thin | — | — | ✓ |
| Scoped, revocable user and org keys | — | ~ | ✓ |
| Decay and freshness tracking (Weibull-derived) | — | — | ✓ |
| Executable steps, not text chunks | — | — | ✓ |
| Guided first install and MCP verification | — | — | ✓ |
| Explicit skill suppression (cognition_forget) | — | — | ✓ |
| Org-level knowledge analytics | — | ~ | ✓ |
| Bandit-optimized retrieval (Thompson sampling) | — | — | ✓ |
under the hood
Want the retention model and neuroscience framing? Keep the homepage focused on setup and proof, then send technical readers to Science.
Start with one verified agent install. Capture one reusable skill. Then make every future run show its source and outcome.
