Skip to main content
Every expert has its own memory. Personal memory is about “who you are”; expert memory is about “what this expert learned while working with you” — your project details, the approaches you discussed, the decisions you made, the traps you hit.

​
How the three layers flow

Memory is a one-way pipeline, from raw to refined:
Core memory is a read-time projection, not a second writable copy. If a fact in core memory is wrong, go back to the memory list and disable or delete the underlying entry — the next consolidation re-distills. Editing core memory directly isn’t a thing.
The core memory and memory graph tabs are behind a platform flag (off by default). When it’s off, only the memory list is available and the other two tabs don’t render — that’s configuration, not a loading failure.

​
How memories are created

​
Conversation extraction

After a conversation with an expert ends, extraction runs in the background. Two conditions must both hold: The counter resets only after a successful write, so a failed extraction doesn’t burn the trigger. Extraction applies the same three criteria as personal memory (specific, durable, non-obvious) and deduplicates before writing.
The most direct route is to say it in the conversation: “Remember: this project’s API must stay v2-compatible.” Explicit memories bypass quality gating.

​
Consolidation (Dream)

Consolidation distills scattered entries into structured knowledge, at two cadences: Deep consolidation only processes memories with activity in the last 24 hours. Each tick has a 600-second budget, and distillation is chunked — at most 12 evidence items per LLM call, committed and marked processed batch by batch. A large backlog therefore won’t time out and lose everything; an interrupted run resumes. The evidence injected into the distillation prompt is bounded: at most 150 entities, 300 facts, and each evidence snapshot truncated to 2,000 characters.
1

Collect evidence

Memory list entries are synced into immutable evidence snapshots, each pointing at its source.
2

Distill in chunks

One LLM call per ≤12 evidence items, identifying entities, facts, and relations.
3

Resolve contradictions

When new information contradicts an old fact (you switched stacks, say), the old fact’s validity window closes and the new one takes over the present. History is never deleted — the full evolution stays on record.
4

Update projections

Results flow into the core memory and graph views.
Consolidation never rewrites the memory list. The entries you manage will not vanish or mutate because a background job ran — that’s a hard design constraint.

​
How much gets injected per conversation

Experts don’t load all memories into context. Each conversation retrieves against a budget: Whichever budget hits first stops the process: 20 items or 2,000 tokens. If retrieval fails, it degrades to a deterministic top-20 — it never falls back to injecting everything. The memories used in an answer are persisted alongside the message, so citations survive a page refresh.

​
Managing memories

Delete and disable are strictly separate: a deleted memory does not come back as a disabled one. When something is merely wrong, prefer editing over deleting — the next consolidation re-distills from the corrected text.

​
Failures and fixes

​
What memory buys you

  • Continuity: no re-introducing project background every time
  • Consistency: it remembers the approaches and design decisions you confirmed
  • No repeats: it remembers the traps you hit and how you got out
  • Fit: it knows your code style and quality bar

​
Creators and memory

Creators can pre-seed an expert’s understanding two ways:
  • Skills: write skill documents that inject domain method
  • Distillation: inject core understanding through the creation conversation
Creator-injected content and user-generated memories coexist in the same system, but user memories are isolated per user × expert — creators cannot see your memory content.

​
Keep reading

Personal memory

Your cross-expert profile

Temporal memory graph

Entities, facts, and validity windows in full

Self-evolution

Background review and skill improvement

Temporary chat

Conversations that write no memory