Memory overview
The canonical version always lives at https://tavina.ai/docs/friendly/memory-skills-habits/memory-overview
Memory overview
Memory helps Tavina carry useful context across conversations. It can help Tavina remember preferences, recurring facts, and working context that would be inconvenient to repeat every time.
Memory is helpful, but it is not a perfect record and it should not be the only source for important decisions. Tavina’s nightly Dreaming refresh can also notice possible memory patterns outside an active chat, but those pattern candidates need review before Tavina treats them as active memory.
What memory is for
Memory can help with things like:
- preferred tone, formatting, or workflow habits;
- recurring names, projects, or context you choose to share;
- corrections you give Tavina;
- information that makes future conversations more useful.
What memory is not for
Do not use memory as the only place to keep:
- passwords, private keys, one-time codes, or payment card details;
- legal, medical, financial, or safety-critical records;
- official account or billing commitments;
- information you do not have permission to store or use.
Review and correct memory
If Tavina recalls something incorrectly, correct it. For important work, ask Tavina what it is assuming and verify the answer against the source of truth.
If Tavina seems to use stale or unwanted context, check Settings and contact support if you cannot find the expected control.
Memory and the Knowledge Library are different
Memory is about Tavina remembering useful context. The Knowledge Library is for document-like Sources, Catalogs, chunks, citations, and long-document workflows.
A document upload should not be treated as the same thing as a personal memory.
Dreaming and pattern review
Dreaming is Tavina’s nightly memory refresh. It runs outside the live chat turn so Tavina can look for reviewable patterns without slowing a conversation down. Each pass covers one full local day for the subscriber, so late-evening conversations are counted on the correct day even though processing finishes after midnight.
Dreaming may flag:
- possible contradictions between remembered preferences or facts;
- open loops, such as unresolved questions, unfinished tasks, or dangling commitments.
These are not automatically treated as settled facts. Tavina keeps Dreaming pattern candidates quarantined in a review-only state until they are confirmed, rejected, kept for later review, or expire. When the Dreaming Review inbox is available in Settings, it separates contradiction and open-loop candidates so you can decide what should become useful memory and what should be ignored. The inbox pages through your full review backlog, so older pending candidates stay reachable rather than being hidden by newer entries, and already-reviewed entries remain visible but clearly marked. If the review list cannot be loaded, the inbox shows a load error with a retry option instead of presenting the queue as empty, and a pattern that has been confirmed and promoted is marked as promoted. Normal memory recall and export should not treat quarantined candidates as accepted memory unless you intentionally review and confirm them.
Memories surfaced in conversations
During a conversation, Tavina can quietly check whether any remembered context would make its next answer more useful โ for example a standing preference or a correction you gave earlier. When a memory is clearly relevant, Tavina may bring a small, relevant reminder into the conversation instead of waiting for you to repeat yourself.
Surfacing stays selective: it only offers a memory when it is relevant to what you are talking about, respects any breaks you take between topics, and never treats a surfaced memory as settled fact โ you can still correct it or tell Tavina to forget it. New accounts now surface useful memories somewhat more readily than before, and per-account settings continue to override the defaults.
You can explicitly ask which memories Tavina used โ for example, “which memories did you use?” โ and it will show the transparent citation list. That list does not appear on ordinary turns; memory use stays silent unless you ask for it.
If you see a surfaced memory you find helpful or unhelpful, accepting or dismissing it tells Tavina whether that kind of reminder was useful, in line with how memory review works everywhere else.
After a turn where a memory was surfaced, you may also see a brief toast asking: “Help train your agent: was this memory relevant to the conversation?” A thumbs-up or thumbs-down answer helps Tavina learn which memories are worth surfacing. The toast fades on its own after a few moments, and letting it fade quietly records nothing โ only an explicit answer counts. If you would rather not be asked, the toast’s “stop asking me again” option turns these prompts off for your account; you can turn them back on any time from the Settings memory controls. The choice is stored with your account, so it applies across your devices.
How Tavina searches memory when you ask
When you ask a remember-style question โ such as what Tavina knows about a person or a preference โ Tavina checks its durable stored memories first before looking anywhere else, so settled facts are not crowded out by looser context.
The search itself is combined: one memory search now covers all of the places Tavina keeps remembered context at the same time, including your profile and memory index, and returns one merged, ranked list. A memory is therefore less likely to be missed just because it lives on a different memory surface.
Tavina can also capture the names you actually use โ nicknames, short names, and blended family or household names โ when saving a memory, and it can find those memories again later using any of those names. If a first search comes back empty, Tavina can retry by breaking a complicated question into smaller, simpler searches instead of giving up.
Evidence checks and confirmations before memory or schedule changes
Tavina double-checks evidence before treating something as an established memory. Important claims about you need support in what you actually said or did โ verified on the server, not just inferred โ and extraction stays deliberately conservative: only concrete, checkable facts are promoted, while vague impressions are left out.
If you challenge something Tavina said it remembers, Tavina verifies the claim against the record before defending it, and concedes when the evidence is not there.
Commitments on your schedule get an extra safeguard. Before creating a scheduled commitment such as an appointment or reminder, Tavina needs verified evidence for it and a single explicit confirmation from you; it will ask rather than silently book something.
These checks make memory slower to jump to conclusions and easier to trust, but memory is still not a perfect record โ keep reviewing what matters.
Memory Control Center
Memory is reviewed from Settings, not from a standalone toolbar tool. The current Memory Control Center is designed to let you inspect what Tavina remembers, search direct and inferred memories, correct mistakes, quarantine suspicious entries, review distillation candidates, and export your data.
Common controls include:
- search memories by content, key, or source;
- filter by memory category;
- distinguish direct memories, inferred memories, and quarantined Dreaming pattern candidates;
- review possible Dreaming contradictions and open loops in the Dreaming Review inbox when it is available;
- mark an entry no longer true, update it, quarantine it, or move it to Recently Deleted;
- review distillation candidates before they become accepted memory.
A deleted memory may remain recoverable for a short Recently Deleted window before final removal. Do not use memory as the only copy of information you cannot afford to lose.
Memory layers shown in Settings
Tavina may show memory and conversation information as related but separate layers. Names and controls can vary by release, but the safe model is:
| Layer | What it means | What to expect |
|---|---|---|
| Conversations | The chat or activity record Tavina can use as source context when it is still retained and available. | It is source material, not a polished memory. Settings may show conversation status, retained-row counts, date range, or storage-size information when available. |
| Daily summaries | Structured summaries Tavina creates from a day of conversation activity. | They help chronology and review. They are not a full transcript dump and may use deterministic fallback wording if full synthesis is not available. |
| Dreaming candidates | Nightly review-only pattern candidates that may point to possible contradictions or unresolved open loops. | They are not settled facts. They should stay review-only until confirmed or expired, and confirmed patterns should behave like regular memories. |
| Semantic memories | Distilled facts, preferences, corrections, and working context Tavina may recall later. | They should avoid copying whole conversations. When source evidence still exists, Tavina may show where or why the memory was remembered. |
MEMORY.md index |
A bounded hot index that helps Tavina find important memories or commitments quickly. | It is an index, not a second profile, transcript, or daily journal. Settings may show index health, entry count or budget, and stale-pointer warnings. |
USER.md profile |
Durable profile and preference context about who the human is and how they prefer to work. | It owns stable profile-style context. It should not be duplicated into the managed MEMORY.md index. |
Provenance, correction, and deletion expectations
When Tavina shows memory provenance, source navigation, or “why remembered” information, treat it as helpful context rather than a guarantee that every original source is still available. A source may be unavailable because it expired, was deleted, was never captured in a navigable form, or is outside the current account/session view.
Correction, forget, and delete actions can have different scopes:
- a correction may supersede a memory without deleting every source conversation;
- forgetting everywhere may remove or retire derived memory/index entries when supported;
- deleting a conversation source may not automatically erase every already-derived summary or memory unless the product says that action includes derived data;
- export or preview controls may show what layers are affected before a destructive action is confirmed.
If the visible Settings controls do not match what you need, contact support instead of assuming Tavina can safely perform a deeper deletion from chat alone.
Memory index health
The memory index health area, when available, is a status view for the managed MEMORY.md index. It may show information such as last curation time, entry count, budget, stale-pointer warnings, or error state. It is intentionally compact and should not expose private internal IDs or raw memory content.