Recall and freshness
Remember starts with recall. It looks up evidence relevant to the prompt and the active project, then adds a compact context block. The agent can ask for more through the memory tools.
Before indexing finishes
A recently captured correction can still be in the retain queue. Automatic recall can search recent queued writes in the same bank and surface a matching original user turn while the background worker indexes it. If matching evidence cannot be included, the hook can flag that memory is pending and try again on a later prompt. Tag-filtered recall skips this queue check, and source text is withheld when source recall or document text storage is disabled.
Automatic recall
- The shared prompt hook defaults to
autoInject: "recall". Git worktree routing keeps the lookup within the project bank unless you explicitly configured a different scope. - The injected block is capped at 1,000 tokens, including its markers, evidence, and any Pages guidance included in the same prompt. It keeps whole evidence items and can omit items that do not fit. Running a manual tool is separate from this automatic budget.
- Recent relevant original user turns can take priority over summaries. The strongest available evidence still depends on what was captured and matched.
remember_reflectis available on demand when you need synthesis.
Handling conflicts
If a user said “use billing-v1” and later corrected it to “use billing-v2,” ask the agent to find the latest correction and to show both values if the history matters. An earlier summary can lag a correction. A timestamp on a turn tells you when something was said, which may differ from the date of the event itself.
Ask for the original wording if a date, identifier, label, approval, or exact edit is important. If the evidence is missing or conflicting, have the agent say what it could establish and leave the rest unknown. Memory cannot grant approval to make a change.