Review and Configure Memory¶
Numi maintains both explicit saved facts and an automatic set of structured memory entries for your account. Relevant memories are appended with new user messages so the assistant retains context across conversations without you having to repeat your background or preferences.
Inspect and edit your memory¶
- Open your profile menu at the bottom of the sidebar and select Settings.
- Open Memory. The panel starts on the Entries tab, organized into four categories:
- You (
ProfileandPreferences) - Areas (long-running projects and systems)
- Topics (recurring subjects of interest)
- People (collaborators and contacts)
- Click any entry to view its dated details. You can edit or delete individual details, or click Add detail to write your own.
- Use Edit summary to change an entry's one-line description, Delete entry to remove the whole entry, or New entry to create one manually.
- Use the search box or category chips (
You,Areas,Topics,People) to filter entries instantaneously in the browser.
Any detail you add or edit by hand is marked yours (manual): it is
authoritative, never expires, and will never be overwritten by background jobs.
Automatically extracted details display their durability lifetime (permanent,
months, or weeks) and when they were last confirmed; details synthesized
across multiple chats are marked cross-chat.
The header bar also shows your Prompt budget meter, which tracks how much of
the MEMORY_USER_PROFILE_TOKEN_BUDGET allowance the always-injected portion of
your memory uses.
| View tab | What it shows |
|---|---|
| Entries | Active and stale memory entries, searchable by keyword and category |
| Pending | Observations extracted from idle chats that are waiting to be consolidated into entries |
| History | Snapshots of earlier memory revisions, newest first |
| Import | Tool for importing exported memories from another AI assistant |
Automatic extraction never records an instruction to agree with you, flatter you, or hold back criticism as a preference. Preferences about tone, length, and format are recorded as usual.
Import memory from another assistant¶
In Settings -> Memory, switch to the Import tab:
- Copy the export prompt provided on the page and run it in your other assistant.
- Paste the resulting list of facts (using headings such as
## Identity,## Projects,## Preferencesand lines formatted as[YYYY-MM-DD] - Fact.or[unknown] - Fact.). - Click Review memories. The memory model organizes the facts into structured entries and shows a preview.
- Inspect the proposed entries and click Save reviewed memories to commit them.
Maintenance actions¶
Click the gear icon next to the view tabs in Settings -> Memory for manual maintenance actions:
- Consolidate now: Immediately folds pending observations into their target entries.
- Cross-chat pass: Re-reads your chat summaries and refreshes the aggregated
details in
ProfileandPreferences. - Export memory: Downloads a Markdown file containing your memory entries, dated details, saved facts, and custom instructions directly from SQLite.
- Clear all memory: Deletes all stored memory and advances your privacy boundary (see below).
To inspect or remove explicit facts saved via the remember_fact tool, open Saved Preferences in Settings. For standing response instructions, see Customize Instructions and Prompts.
What reaches the model¶
Each new user message can carry a stored, model-only memory_note containing
personal instructions, profile details and relevant memories. The first note
establishes context; later notes append only new details, corrections and
withdrawals. Previously stored notes are replayed unchanged, including after
reloading a chat. Notes and their bookkeeping are excluded from browser message
payloads, account chat exports and public shares.
To see what the model received, open Chat info from a chat's menu in the sidebar. Memory used lists your instructions, each subject with the exact details sent, saved preferences, the entry list and recent chat titles, and the message that added each one. Expand a row to read its lines. Chats from before stored notes show an estimate, and nothing is listed after you turn memory off or clear it.
While the past_chats tool is available, the note also lists the titles and
start dates of your 10 newest other chats. Temporary and untitled chats, and
chats from before a memory reset, are left out. The model sees titles only and
opens a chat with past_chats when your message seems to continue one. The
list is ordered by start, so a later note repeats it only after you start or
delete a chat. As with other memory corrections, earlier notes stay unchanged;
the newer list overrides them, and losing access to past_chats withdraws it.
The current message selects up to MEMORY_ENTRY_EXPAND_LIMIT subjects. With
semantic search enabled and the index ready, multilingual embeddings match a
message in any language against the English memories. Only subjects that
clearly stand out from your other subjects for that message are added, so
general questions add little or nothing. Otherwise Numi uses the existing
name/summary word search. Either way a compact list of subject names is
included, and the model can call recall_memory for details that were not
added; an explicit recall always returns the nearest subjects. Retrieval never
renews expiry. Automatic profile context (Profile and Preferences) is
limited to half of the note budget so it cannot crowd out current search hits.
Matching subjects receive details in rounds;
one large subject cannot consume the budget before the others get a place.
The subject list and recent chat titles get their room before any details, so a
full note never drops them; a list longer than the whole budget is left out.
MEMORY_USER_PROFILE_TOKEN_BUDGET bounds each new context addition, including
explicit facts; their individual PROMPT_MAX_MEMORY_* limits also apply.
Unchanged facts are not appended again. This avoids rebuilding the opening
memory prefix whenever memory changes, but does not guarantee fewer input
tokens for every conversation: useful new details still grow the history.
Existing chats transition once when their first stored note is added; their old dynamically generated prefix was never recorded and cannot be recovered byte for byte. Context-window truncation/compaction can rebuild a bounded context snapshot. Disabling account memory or clearing it excludes earlier notes from future requests for privacy, even though that invalidates a cache.
Operator: optional semantic search and daily housekeeping¶
Both additions are opt-in and independently disableable. Switch them on the
admin Memory page (Recall and Upkeep) without a restart, or set the
starting values in .env:
MEMORY_EMBEDDINGS_ENABLED=true
MEMORY_EMBEDDING_MODEL=openai/text-embedding-3-large
MEMORY_MAINTENANCE_ENABLED=true
Semantic search uses your configured OpenRouter credentials and sends the account's current subject names, summaries and details, plus search text, to OpenRouter. It adds no vector database service: rebuildable float32 vectors live in SQLite and are scoped through the owning memory entry. The background worker indexes up to 64 changed subjects per user per five-minute sweep. Edits, expiry, reset and deletion invalidate old vectors immediately; an incomplete index, missing credentials or provider failure uses word search. Failures are logged. Each message adds one short embedding request, roughly half a second with the default model, before the answer starts.
Relevance is judged relative to the account's other subjects instead of by a
fixed similarity score, so another embedding model needs no recalibration;
changing the model only rebuilds the index. The default is served by OpenAI
and Azure through OpenRouter. In a test with German questions against English
memories it found the right subject for 47 of 48 questions; baai/bge-m3
and qwen/qwen3-embedding-8b are cheaper, but BGE-M3 missed about a third
of them and both are served by third-party hosts.
Daily housekeeping runs an agent at medium reasoning effort, at most once per
account per 24 hours, and not again until memory changes. Its model is
MEMORY_UPKEEP_MODEL (DeepSeek V4.1 Flash by default; inherit uses the
memory model), and the admin Models page can override it as Daily memory
upkeep. The agent works through the account's memory with tools: it
lists and reads subjects, finds likely duplicates, merges subjects that are the
same, moves details to the subject they belong to, retires details without a
lasting personal fact (a one-off question, a duplicate, a contradicted point),
rewords details, fixes names and categories, and writes a short summary for
each subject that says what it covers. Those summaries are what the chat model
sees in the subject list. It also tidies the automatic Profile and Preferences
details. It cannot create new facts; extraction still handles new
conversations. Manual and imported details and subjects, and the Profile points
rebuilt by weekly aggregation, are protected. Every change is checked and saved
on its own, so one rejected edit does not discard the others, and a subject a
chat changed meanwhile is refused until the agent reads it again. There is no
limit on steps or output length. A review that runs for 10 minutes, or keeps
repeating the same call, ends for the day and keeps its committed edits, so one
account cannot hold up the memory worker. A tidy memory costs a few short
requests. When a review changes something, before/after snapshots appear
in Memory history, which lists what was added, removed, reworded, moved or
renamed in each version. Failed runs wait an hour before retrying.
Housekeeping never changes a detail's first-seen, last-confirmed, durability or expiry fields. Consolidation dates facts by their original user message; cross-chat aggregation cites supporting chats and uses their evidence dates. Job execution timestamps are tracked separately, so rereading old data cannot extend its lifetime merely because a background job ran again.
Memory jobs do not cap the model's output length. Extraction uses medium
reasoning effort; the other structured jobs use low effort. They do not reread
every raw conversation daily.
The model's judgments can still be wrong: inspect the first housekeeping
snapshots before relying on automatic organization. Turning either setting
back to false stops that feature without removing original memory or chats.
Clear stored memory¶
In Settings -> Memory, open the gear menu and select Clear all memory.
This permanently deletes all entries, details, explicit saved facts, pending
observations, chat summaries, and memory history for your account, while keeping
your chats and custom instructions intact. It also sets a timestamp boundary
(memory_cleared_at) on your account so chats from before the reset are never
re-extracted.
- Opting out completely: Clearing memory wipes current data, but new chats
can still build new memories. To disable memory use and collection across your
account, turn off Use memory in Settings -> Privacy & data (this also
disables the
past_chatstool). - One-off private chats: Enable Temporary at the top right of a new chat when you want a conversation that never contributes to memory, search, or the file library and is deleted automatically when you leave (or after 24 hours of inactivity).
Operator: choose the memory model¶
Background extraction, consolidation, and cross-chat aggregation use a dedicated lightweight model so they don't consume your main chat model's quota.
On the admin Models page, find Memory extraction and summaries, pick any
enabled catalog model, and click Save override (or Use config to revert
to MEMORY_SUMMARY_MODEL in .env).
Operator: rebuild memory from existing chat history¶
Because Clear all memory moves a user's privacy boundary forward to block re-reading old chats, operators have a dedicated CLI tool when they want to discard derived memory and re-extract it from existing chat history:
uv run memory-rebuild --user <name> # dry run (shows eligible chat count)
uv run memory-rebuild --user <name> --apply # rebuild from full history
uv run memory-rebuild --user <name> --apply --since 2026-09-06
memory-rebuild removes derived entries, details, observations, chat summaries,
and snapshots while preserving hand-written (manual) details unless
--drop-manual is passed, and moves the user's boundary back to --since (or
clears it). The background worker then drains the eligible chats oldest-first in
bounded batches (MEMORY_EXTRACTION_IDLE_BATCH_SIZE per five-minute sweep).
Operator: tune or disable the pipeline¶
Set these MEMORY_* variables in .env and restart the server. The ones
listed under runtime settings
can also be changed on the admin Memory page while Numi runs:
| Variable | Default | Effect |
|---|---|---|
MEMORY_ENABLED |
true |
Enables automatic background extraction, consolidation, and aggregation |
MEMORY_SUMMARY_MODEL |
deepseek/deepseek-v4.1-flash-1017 |
Default model for memory jobs when no admin UI override is set |
MEMORY_EXTRACTION_COOLDOWN_MINUTES |
90 |
Idle minutes after a chat's last message before extraction runs |
MEMORY_EXTRACTION_CONSOLIDATION_THRESHOLD |
5 |
Pending observations required to trigger consolidation |
MEMORY_EXTRACTION_IDLE_BATCH_SIZE |
4 |
Maximum chats and consolidations processed per five-minute sweep |
MEMORY_EXTRACTION_IDLE_FLUSH_MAX_AGE_HOURS |
0 |
Maximum age of idle chats eligible for extraction (0 = no age limit) |
MEMORY_AGGREGATION_ENABLED |
true |
Enables the periodic cross-chat pass for Profile and Preferences |
MEMORY_AGGREGATION_INTERVAL_HOURS |
168 |
Minimum hours between cross-chat aggregation passes (7 days) |
MEMORY_MEMORY_LANGUAGE |
English |
Language used to store internal memory entries and summaries |
MEMORY_EXCLUDE_SENSITIVE |
true |
Instructs extraction to skip health, religious/political beliefs, identity, and financial details |
MEMORY_SLOW_TTL_DAYS / MEMORY_SEASONAL_TTL_DAYS |
180 / 21 |
Days before slow and seasonal details decay to stale |
MEMORY_USER_PROFILE_TOKEN_BUDGET |
2500 |
Estimated token budget per appended memory note |
MEMORY_EMBEDDINGS_ENABLED |
false |
Enable optional multilingual search through OpenRouter |
MEMORY_EMBEDDING_MODEL |
openai/text-embedding-3-large |
Embedding model; changing it rebuilds the index |
MEMORY_MAINTENANCE_ENABLED |
false |
Enable daily organization of existing memory |
MEMORY_UPKEEP_MODEL |
deepseek/deepseek-v4.1-flash-1017 |
Model for the daily upkeep agent; inherit uses the memory model |
Notable pipeline behaviors:
- Resumed chats: If you continue an older chat after it was already extracted, the extractor re-reads the updated conversation in full and supersedes its earlier pending observations.
- Zero-request fast paths: Consolidation confirms near-duplicate details in place and populates brand-new entries directly without making an extra LLM call.
- Graceful decay: When a
sloworseasonaldetail passes its TTL, it is markedstalerather than deleted. Stale details drop out of automatic prompt injection and normal recall. Historical snapshots retain earlier revisions; stale details are not silently reconfirmed by a lookup or housekeeping run.
See Adaptive memory settings for the complete list of MEMORY_* settings.
Operator: why stored memory uses a single language¶
Numi writes internal memory entries in a single configured language
(MEMORY_MEMORY_LANGUAGE=English by default) regardless of the language used in
individual chats. Keeping the memory store in one language allows deduplication
and similarity checks to match facts across multilingual conversations, while
the assistant still replies in whatever language you use in the chat.
Storage and API access¶
Extraction reads only user questions and final assistant answers (never tool outputs, reasoning traces, or system prompts). Custom clients can inspect and manage memory through the authenticated memory endpoints.