Features

LivingMemory is not just a chat log. It transforms conversations into searchable, decay-aware long-term memories that agents can also use directly.

Automatic reflection

After the configured number of rounds, the plugin asks the LLM to summarize recent conversation into structured memory.

Rich retrieval content

Retrieval content is assembled from summary + key_facts to preserve information density. persona_summary powers memory injection, dashboard display, and persona expression.

Agent memory tools

Agents can call recall_long_term_memory to search old memory or memorize_long_term_memory to store durable facts.

Memory atomization

Important facts become independent atoms with type, TTL, importance, access count, and decay state.

Where memories come from

Path	Trigger	Best for
Automatic reflection	Conversation reaches the configured summary rounds	Long-term preferences, project context, relationships, durable facts
Agent write tool	The model calls memorize_long_term_memory	Explicit "remember this" requests, important agreements, long-running tasks
Manual administrator summary	/lmem summarize [message_count]	Save the current context immediately or re-summarize the latest N messages

Summarization produces a persona-styled summary used for both memory injection and dashboard display (persona_summary), while retrieval content is assembled by code from summary and key_facts so a single compressed summary cannot drain information density. Custom prompts may still emit a factual canonical_summary for consumers such as graph extraction. Memories above the source-retention threshold also store their source messages separately in SQLite. Those messages are not indexed for normal recall, but remain available for Dashboard review, re-summarization, and deep source retrieval by the Agent recall tool.

How recall works

Before the LLM request is sent, LivingMemory retrieves relevant memories. Results can be appended to the user message, placed before or after it, or injected as simulated tool-call context.

Ranking combines:

Factor	What it does
Keyword match	BM25 and graph keyword retrieval quickly find concrete entities and phrases
Semantic similarity	Vector retrieval handles different wording with similar meaning
Graph relationships	Entities and cross-memory edges add structure across facts
Importance	Durable preferences and agreements are favored
Time decay	Old memories gradually lose weight unless repeatedly accessed or reinforced
Retrieval boundaries	Candidates can be filtered by minimum importance, vector similarity, and memory type
Recent-memory reserve	Dedicated slots keep recent in-scope memories from being displaced by similarity ranking alone
Recent context	Recall can include recent conversation within a configured age limit to reduce topic drift

Opt-in agentic recall (beta)

agent_tools.enable_agentic_recall_beta (off by default; requires agent_tools.enable_recall_tool) extends active recall into a multi-turn loop within one request:

Capability	Behavior
People-centric search	target="current_speaker" searches the current speaker's experiences; start_time/end_time filter by event time; exclude_ids and search_offset support cursored paging
Evidence reading	read_memory_evidence returns a memory's active facts and, when the whole source is verifiably visible to the session, its retained text; facts and source are cursor-addressable so a truncated read resumes exactly
Shared budget	Tool calls, cumulative tool time and per-call output share one per-request budget; when it is spent the model answers from the evidence it already gathered
Add-only rules	The recall rule prompt is injected as a temporary part, never into conversation history or the plugin's message store

With the switch off, tool registration, retrieval and prompts are identical to the default behavior.

Who can access which memories?

The same boundaries apply to both memory writes and recall:

Capability	Behavior
Memory scopes	session isolates each conversation, user shares across private and group chats for one platform user, global shares broadly, and legacy preserves earlier behavior
Forced isolation	Selected conversations always keep an independent scope, even when a shared mode is active
Memory allowlist	Limits capture, summarization, recall, and Agent memory tools to selected users, groups, or complete session IDs
Identity aliases	Maps platform identities to stable names before summarization; graph participants also reuse stable identities instead of splitting on nickname changes

Scope changes affect newly written memories only. Existing data is not migrated automatically.

Lifecycle behavior

Mechanism	Purpose
Importance decay	Reduces the weight of old low-value memories
Access reinforcement	Frequently recalled memories are more likely to stay relevant
Atom TTL	Different fact types can age differently
Cleanup or archiving	Low-value memory can be deleted permanently or removed from recall indexes while retained as a recoverable archive
Important-memory protection	Memories above a configured threshold are excluded from daily importance decay
Source retention	High-importance memories can retain source messages for audit and re-summarization
Safe backups	Creates backups before version updates and migrations

Archived memories do not participate in normal recall. Restoring one from the Dashboard regenerates its embedding and rebuilds BM25, graph, atom, and other derived indexes.

How large index maintenance stays available

After startup, LivingMemory checks document, vector, BM25, and graph indexes in the background instead of blocking plugin readiness when extensive repairs are required:

Mechanism	Purpose
Provider fingerprint	A changed embedding provider or model triggers a full vector generation rebuild, even when dimensions are unchanged
Batching and checkpoints	Read size, embedding batches, concurrency, and delays are bounded; full vector rebuilds can continue from checkpoints
Shadow indexes	BM25, FAISS, and graph data are built in a separate generation and switched only after success; failures leave the live generation intact
Concurrent-write reconciliation	A final consistency pass repairs data written or changed while maintenance was running
Visible status	/lmem status reports checking, rebuilding, partial, failed, and cancelled states with progress

See WebUI management for data operations and Configuration for all related settings.