Memory
Elephant Agent memory is not one bucket. It is a layered system that separates durable understanding from runtime trace, current-turn support, and background learning.
Recall is part of memory, but memory is not reduced to recall. A retrieved Step can support the current turn; it becomes durable understanding only when a governed update writes through the Personal Model.
Memory layers
| Layer | What it remembers | How to inspect it |
|---|---|---|
| Personal Model | Active, retired, and disputed claims across Identity, World, Pulse, Journey. | Dashboard Personal Model, Diary, and source views. |
| Elephant State | Elephant identity and one current context note. | Herd, status, wake resume. |
| Episode / Loop / Step trail | Inputs, replies, tool calls, results, updates, and provenance. | History, runtime trace, Why inspection. |
| Contextual recall | Current-turn support retrieved from facts and Steps. | /recall, diagnostics, search result status. |
| Background learning | Episode-close, diary, dream, questions, and visible skill-matching updates. | Reflect, jobs, learning results. |
Claim-aware recall
Personal Model search returns claims with match status. It does not pretend that every old note is support.
tool.personal_model.search can combine:
- topic, claim, and source-support field matches
- Unicode lexical matching and CJK n-grams
- weak fuzzy matching for small spelling or character variation
- semantic retrieval from the durable index
- optional translated or paraphrased
query_variants - lightweight polarity checks for clear preference statements
| Status | Meaning | How Elephant Agent should behave |
|---|---|---|
strong_match | The Personal Model contains reliable support. | Use the claim with appropriate confidence. |
weak_match | The result is relevant but not strong enough to overstate. | Mention uncertainty or ask if needed. |
no_match | There is no reliable Personal Model support. | Do not fabricate memory. |
The public modes are intentionally small: use mode=auto for claim lookup and
mode=inventory for the lens/topic map. Exactness comes from topic, ref,
status, and diagnostics; cross-language or conceptual lookup can use
query_variants.
Current-turn recall
When a current turn needs past support, Elephant Agent may inject a
Current-turn recall support: block.
| It can do | It cannot do |
|---|---|
| Bring relevant evidence into this turn. | Become Personal Model truth by itself. |
| Explain why a reply is grounded. | Override an active corrected claim. |
Return no_match honestly. | Invent a memory when support is weak. |
Current-turn recall support is useful evidence. It is not a hidden permanent memory write.
Inspect understanding in practice
Use the dashboard Personal Model page to inspect and correct claims, Curiosity
to manage useful prompts, and source/history views to trace the support material
behind them.
Diagnostics can explain why a search matched or returned no_match; they are for
inspection and debugging, not prompt truth.
Related pages
- Embeddings for local semantic retrieval and provider posture.
- Correctable understanding for remember, correct, forget, and dispute.
- Background learning for episode close, diary, dream, and reflect jobs.