Assistant memory
The Memory tab on an assistant's detail page gives owners and admins a live view into that assistant's cognitive memory vault — the three-tier store that Chiro draws on when answering questions, recalling decisions, and building its sense of your organization.
Access: navigate to the assistant's detail page at /{org-slug}.achiral.ai/assistants/{id} and select Memory from the tab bar. The tab is hidden from members without the owner or admin role.
How Chiro Memory works
Chiro Memory uses a proprietary cognitive activation model. Every memory object in the vault carries an activation score derived from:
- Recency — memories accessed recently retain higher activation.
- Frequency — memories retrieved many times are reinforced.
- Decay — unused memories fade at a tier-specific rate.
The three tiers each have a different decay rate:
| Tier | What it stores | Decay |
|---|---|---|
| Episodic | Recent conversations and events | High (fades quickly) |
| Semantic | Decisions, preferences, and domain knowledge | Low |
| Core | Long-term documents and persistent context | Near-zero |
A nightly consolidation job promotes episodic memories that remain salient into the semantic tier, mirroring the way the human brain consolidates memories during sleep.
Memory tier distribution
The first card on the Memory tab shows how the assistant's memories are spread across the three tiers.
- A stacked progress bar represents the proportional share of each tier.
- The legend beneath it shows the exact count and a short description of each tier's contents.
- The total memory count appears in the top-right corner of the card.
This distribution is fetched via GET /api/assistants/{id}/memory.
Top activated memories
The second card lists the five memories with the highest current activation scores.
Each row shows:
- Tier badge — episodic, semantic, or core.
- Snippet — a short preview of the memory content.
- Collection — the Weaviate collection the object belongs to.
- Last accessed date and retrieval count.
- Activation bar — a visual indicator and percentage representing the computed activation score (0–100%).
Higher activation means the memory is more likely to surface during inference without an explicit semantic match.
Core memory candidates
The third card is the flashbulb candidate queue — memories that the salience scorer flagged as high-importance and are waiting for human validation before being promoted to permanent core memory.
A memory becomes a candidate when its multi-signal salience score (a combination of emotional intensity, explicit intent, and semantic novelty) crosses the configured threshold.
When the queue is empty, a green confirmation notice appears. When candidates are present, an orange notice explains the decision:
"These memories scored above the salience threshold and are candidates for permanent core memory. Approve to retain them indefinitely; reject to remove them from the knowledge base."
Each candidate row shows:
- A content snippet (up to three lines).
- Its current tier, collection, and salience score (as a percentage).
- Two action buttons: Approve and Reject.
Approving a candidate
Clicking Approve calls PUT /api/assistants/{id}/memory/candidates/{weaviateId} with { "action": "approve" }. The memory is promoted to the core tier and removed from the candidate queue immediately in the UI.
Rejecting a candidate
Clicking Reject calls the same endpoint with { "action": "reject" }. The memory object is deleted from the Weaviate tenant and removed from the queue.
Both actions require the owner or admin role. The button shows a spinner while the request is in flight; the row disappears on success.
FAQ
Who can see the Memory tab?
Only users with the owner or admin role on the organization. The tab is not visible to regular members.
Does rejecting a candidate delete the original conversation? No. It deletes only the Weaviate memory object. The conversation messages in MongoDB are not affected.
How often does the candidate queue update? New candidates appear as Chiro has conversations. The queue is fetched fresh each time you open the Memory tab (lazy-loaded on first visit).
What is the activation score scale? Activation scores are normalized to 0–100% for display. A score of 100% means the memory had the highest activation in the retrieved set.