Cross-session long-term memory

Long-term memory keeps the caller’s profile, expression preferences, stable facts, and durable interests for later sessions. Data is isolated by workspace and caller identity: workspaces, IM users, and embed visitors each maintain personal memory.

Enable and use long-term memory

1. Workspace Admin/Owner: Settings → Long-term memory, enable the switch (off by default).
2. Choose Explicit memory only or Automatic extraction; in auto mode, optional extraction model (empty = conversation model).
3. In personal memory management you can disable memory for yourself; an enabled workspace does not force it on you.
4. In conversation, state clearly what to remember, or add manually via memory management; verify in another session.
5. Review pending confirmation entries regularly: confirm, edit, or reject inferences; delete what is no longer valid.

Mode	Behavior
explicit_only	Records only what is explicitly requested; no background LLM distillation
auto	Additional background extraction from conversations, merged per delay and minimum interval

memory_enabled=false on an agent disables memory read/write for that agent; omitted = inherit workspace settings. IM/Embed follow the linked agent’s memory preferences; the API key must be full-access for personal memory management APIs. The feature depends jointly on workspace, personal profile, and request context.

Manage memory

Action	Usage
Add / edit	Maintain profile, preferences, facts, tasks, or interests; manual edit is no longer overwritten by auto extraction
Confirm pending	Inferred entry goes from pending to usable; unconfirmed = not injected into the prompt. If inference modifies an existing entry, the old one stays active until confirmation then replacement; expired inference → confirmation fails, refresh the list
Reject / delete	Cancel wrong or obsolete memory; rejection leaves an inhibition trace to limit re-extraction
Topic elevation	Promote a recurring subject to durable interest immediately; stop tracking a theme
Document preferences	See often-cited documents; remove a personalized search preference that is no longer useful
Immediate consolidation	Merge close entries, archive due tasks, without waiting for background job
Export / clear	Download your memory JSON, or wipe your memory data

Items cover profile, preferences, facts, tasks, and interests. A new fact on the same topic may replace the old one. Disabling profile suspends use; delete or clear removes data. Categories and statuses in the Memory API.

Impact on answers and retrieval

The system injects profile, preferences, and interests as permanent context (length budget), and recalls facts and tasks relevant to the question. With a configured vector model, semantic recall ranks all identity entries by question relevance, without relative importance; with pgvector on PostgreSQL sorting is in-database, otherwise computed in the service. In intelligent reasoning, active search in memory and history. Memory only affects understanding and context choice, not knowledge base access rights.

With retrieval personalization enabled, themes help interpret the question and often-cited documents rerank. The conversation timeline shows memory-related steps. Background jobs track extraction cursor and pending state per session: short bursts, batch truncation, or restart do not lose messages. If model output stays invalid on a segment, up to three retries then skip the segment without auto re-extraction of the skipped segment.

Troubleshooting: memory not applied

Symptom	Check
Memory not used	Workspace, personal, and agent switches; workspace or identity change
No auto extraction	write_mode, extraction delay/interval, model and background jobs
Inference has no effect on answer	Still pending; usable only after confirmation
Stops citing a document	Remove in document preferences; KB rights remain independent
Immediate consolidation merges nothing	skipped reason in response (few items, no candidates)

Shared memory: team and agent

Besides each person's own memory, a workspace can keep shared knowledge. There are two scopes, each configured separately in "Settings → Long-term memory → Shared memory" (admin):

Scope	Content	Who benefits
Team	Decisions, conventions, customer or project context	Every member of the workspace
Per agent	An agent's own rules and knowledge	Everyone who talks to that agent

For each scope:

Setting	Effect
Off	Nothing is read or stored (default)
On request	Only notes added by an admin, or asked for in a conversation, are stored: "Remember for the team: …", "Remember for this agent: …" (or « Retiens pour l'équipe : … », « Retiens pour cet agent : … »)
Automatic	In addition, background extraction proposes shared notes from conversations
Admin approval	Proposed notes wait for approval before being used; a rejected note is not proposed again
Maximum notes / notes per answer	The scope's capacity and how many relevant notes are added to each question

Shared memory is independent of the personal memory mode: for example, keep personal memory explicit-only while letting the assistant propose team notes automatically. Notes written by an admin are standing rules, added to every answer; other notes are added only when they concern the question. They are shown to the model separately from personal memory. An agent with memory turned off neither reads nor writes shared memory.

"Manage shared notes" lets admins add, edit, approve, reject or delete notes; for the agent scope, pick the agent first. API: /api/v1/memory/shared/{workspace|agent}/items (Admin role, ?ref=<agent id> for the agent scope).

Projects and themes

Two independent options, in "Settings → Long-term memory → Projects and themes" (admin), give the assistant the big picture beyond single notes.

Projects. An admin creates a project (name, description) and may link an agent: every conversation with that agent is linked to the project. The assistant keeps a Markdown project brief (goal, decisions, current status, next steps, open questions), updated after conversations when "Update briefs automatically" is on. The brief can be edited by hand, regenerated from all conversations, and is added to the context of every conversation of the project. An agent feeds only one project; deleting a project keeps its conversations.

Automatic themes. Like Agent Memory's "scenes", the assistant groups the notes of a memory space (each person, and the team when the Team scope is on) into a few themes, each with a short summary. Regrouping happens at most once per configured interval (24 h by default) or on demand ("Regroup now"). Themes related to the question are added to the context. Everyone sees their own themes in their personal memory; admins see the team's.

Setting	Effect
Projects / Update automatically	Turns projects on; updates briefs after conversations
Automatic themes / Maximum / Interval	Turns themes on; caps them per space; minimum time between two regroupings
Model for briefs and themes	Empty: the memory extraction model, otherwise the workspace's

API: /api/v1/memory/projects (read for members, changes for admins), /api/v1/memory/themes (personal themes), /api/v1/memory/shared/workspace/themes (team themes, admin).

Long-term profile

Ported from Agent Memory (layer 3, "persona"), set under "Settings → Long-term memory → Long-term profile" (admin). Two independent options, off by default:

• Profile of each person: at most 2,000 characters, in four layers: base facts, interests, how to answer them, how they decide. It is written from the person's themes and memories, inventing nothing and keeping no sensitive data.
• Team doctrine: at most 1,200 characters. These are the principles, routines, decision rules and boundaries the team keeps to, drawn from team memory. Project-specific facts, versions and progress stay out. It needs team memory to be on.

When the text is written. A first time once the space holds at least 4 memories. Then after each batch of new memories (10 by default, configurable). Before a rewrite, a quick check compares the new memories with the current text. It only rewrites when they contradict it, change the scope of a rule or add a rule. When they only confirm it or add nothing durable, the text stays as it is, with no paid rewrite.

Where to see it. Everyone sees, corrects or erases their profile in "My memory". Admins handle the team doctrine in the settings. A text edited by hand is no longer rewritten automatically, until someone clicks "Regenerate". The profile and the doctrine are added at the top of every conversation's context.

API: GET/PUT/DELETE /api/v1/memory/persona and POST /api/v1/memory/persona/rebuild for one's own profile; the same routes under /api/v1/memory/shared/workspace/persona for the team doctrine (admin).

Workspace configuration

Workspace admin manages switch, extraction mode, and model in settings. Config in memory_config, read/updated via GET/PUT /tenants/kv/memory-config; full fields in the Memory API.

Name	Type	Default	Description
enabled	bool	false	Workspace switch
write_mode	string	explicit_only	explicit_only / auto; other values treated as explicit_only
extract_model_id	string	empty	Extraction model; empty = conversation model
extract_delay_seconds	int	90	Delay after reply before extraction, merges nearby turns; 5–3600
extract_min_interval_seconds	int	300	Minimum interval between two extractions for one person; messages in interval deferred; max 86400
extract_instructions	string	empty	Workspace-specific extraction rules, max 1000 characters
max_items	int	200	Active item cap per identity, max 2000; 0 = default
interest_threshold	int	3	Sessions a theme must appear before durable interest, max 20
embedding_model_id	string	empty	Vector model for memory recall; empty = lexical match only
vector_recall	bool	omitted = on	Semantic recall when vector model configured
retrieval_conditioning	bool	omitted = on	Allow memory to influence question and document ranking

The extraction model is for background synthesis; the vector model for matching with the question. Memory increases model calls; deleting a model referenced by memory config shows dependencies.

Implementation reference

• internal/application/service/memory/: scope, extraction, recall, themes, document preferences, consolidation
• internal/handler/memory.go, internal/router/routes_memory.go: personal API
• internal/types/memory.go: models, budgets, states, and config
• frontend/src/views/settings/MemorySettings.vue, MemoryWorkspaceSettings.vue