The complete memory lifecycle
Build experience that stays useful. Find stronger evidence, revisit what was known, run optional intelligence on your device and make every reusable skill earn its place.
Across turns and agents
Preserve what happened, what matters and what comes next. Give the next turn a starting point grounded in the work already done.
Connect your agent loopCapture selected documents and visible conversations with stable source identities. Direct text, Markdown and JSON retain their exact text; explicit extractors handle images and other binary documents. Selected file watches track revisions.
Prepare context before a turn, capture the visible exchange after it, and queue bounded follow-up work. Explicit event adapters accept selected, completed agent messages. Durable reservations guard against repeating an ambiguously completed turn.
Carry completed work, open tasks, decisions, constraints and rejected approaches across sessions and explicitly shared agents. Source-dependent checkpoints stop being resumable when their evidence becomes invalid.
Define typed preferences or project state. Each field reports a supported value, an unknown, or conflicting candidates with source citations. Reads recheck the source state before returning values.
Explore typed profilesMnemosyne Recall
Select the right level of detail for the task. Keep the original within reach when a summary leaves something out.
Build adaptive contextLocal relevance scoring weighs rare terms, repeated terms and document length across the memories authorized for the query. Records outside the query’s authorized scope do not influence those statistics. Untrusted and time-ineligible records are excluded by default. The earlier overlap scorer remains an explicit compatibility option.
Combine lexical and semantic candidates, then optionally rerank them. Semantic retrieval considers the full authorized indexed scope with bounded candidate memory and a deadline. Vectors retain their model identity and source revision; eligibility is rechecked after provider work.
Run optional embeddings and reranking on your CPU. Provision the supported runtime and model weights explicitly, then use cached models offline. The default local store and lexical recall need no model. Observation and reflection use separately configured providers.
Set up local intelligenceAssemble eligible observations, hierarchical summaries and original excerpts into a bounded prompt. Supply your tokenizer or use the labelled conservative byte count. Building context and reusing its selection starts no model call.
Follow an issued source handle to the original UTF-8 text in bounded pages. Selection caches retain references and ranges. Changed evidence invalidates old handles and generated context.
Connect explicit entities and aliases through cited relationships. Explore a bounded neighborhood while retaining the sources and ownership behind each connection.
Explore retrieval optionsMnemosyne Timeline
Keep two clocks: when something held in the world, and when the system learned it. Make the next turn current without losing the history behind it.
Inspect freshness checksChoose a world-time date and a knowledge-time cutoff for recall and compiled context. Dependency, conflict and outcome checks follow those clocks together. Historical results do not restore erased evidence or make failed outcomes successful.
Explore temporal contextPreserve the history and recorded cause of a correction. With sufficient evidence, prior guidance remains available before the correction’s effective time while obsolete dependent guidance is withheld afterward. Ambiguous legacy invalidations stay withheld; changed sources may conservatively retire more guidance than necessary.
Different active claims sharing an explicit fact key remain a conflict. Profiles can retain competing source-backed candidates. Similar wording alone never silently establishes that two facts are identical.
Set a recheck policy and supply bounded source probes. Confirmed, changed and unavailable results feed freshness-aware recall. A fact's age prompts a check; it does not invent a replacement or increase trust.
Bind an action to its declared memory dependencies. Revalidate revisions, source-check state and outcomes immediately before the host callback. The application still supplies complete dependencies and controls the external action.
Follow a correction in the demoMnemosyne Evidence Gate
Turn selected experience into observations and tested procedures, with evidence that can be inspected, challenged and retired.
Inspect the recorded learning cycleAn explicit proposer turns supplied sources into cited observations. Jobs retain state through restarts, with leases, bounded retries and call, byte and time limits. Start one drain or a finite worker when your application needs it.
Build overview and detail representations from selected evidence. Reuse unchanged generations and refresh when inputs change. Full generation checks include ancestor records, source verification and successful or failed outcomes.
Describe a procedure’s prerequisites, steps, parameters and evidence. We recommend promotion after passed trials across at least two distinct tasks and two distinct verifiers. Candidate skills accumulate trials while remaining outside ordinary advice; duplicate task or evidence submissions do not add promotion credit.
Promotion requirements are saved with each new skill. Reopening the store with weaker options cannot lower them. The compatibility default and legacy skills retain one-task, one-verifier requirements; configure the stronger policy for new candidates. Failed trials or prerequisites retire a skill.
Record which memories were actually used and how the task ended. Bounded reflection proposes lessons, rejects invented citations and checks source revisions before commit. These operations update external memory; the host supplies the outcome evidence.
Explore reflection and validationMnemosyne Migration
Connect an agent where it already works. Start beside an existing memory system or make a deliberate, reviewed import.
Explore gradual migrationA read-only legacy search adapter copies the records your agent retrieves. Policy-gated paired recall checks can move routing from shadow to assist to local preference. The old store is checked on every query and stays unchanged, with immediate legacy-routing rollback. Progress covers observed queries, without assuming the whole store has migrated.
Preview and atomically apply memories through 11 selected export profiles. Inspect original record bytes, replay safely and undo unchanged batches without discarding later work.
See supported export formatsUse the TypeScript SDK, CLI, MCP server, authenticated HTTP service or bundled Python client. Share selected workspace records across agent identities while ownership continues to govern edits.
Connect through MCPExpose scoped virtual memory files through provider memory tools and supported function-calling interfaces. Your host keeps control of model requests, conversation state and permissions.
Read the provider integration guideAn inspectable local system
Own the store and the decisions around it. Review proposed changes, set boundaries and rehearse recovery before relying on memory.
Read the operating boundariesRun the local engine without a separate database server. Records are private by default. The HTTP inspector uses scoped, revocable credentials, with independent capture, recall, read-only and destructive-operation controls.
Stage additions and corrections in a private memory branch. Inspect the proposed change, then merge atomically only while its base evidence remains current. Drafts and controller job records stay out of ordinary advice.
Take a consistent whole-database snapshot, verify its digest and database integrity, then restore to a new path. Preserve workspaces, pending jobs and erasure tombstones. Scoped exports remain a separate transfer option.
Rehearse recoveryErase source generations and dependent live content. Hashed tombstones block replay after restart, including across full and gradual migration when the source identity matches. Legacy stores and existing external copies remain under your own retention controls.
The 500-question matched corpus run reports complete source-session coverage on 425/470 answerable questions at retrieval and 382/470 after 8,192-byte packing. This is evidence coverage, not answer accuracy, with same-day-compatible source timing. The separate agent harness can measure answer quality, stale actions and negative transfer with your own reader.
Inspect the evaluation protocolKeep vector and hybrid recall with improved relevance scoring, broader startup indexing and scoped cache revalidation. The Mnemosyne Vector Engine remains separate from the local engine, with explicit migration when you choose to move records.
Read the existing-user guideBuild from the tagged source, follow the integration guides and run the included examples.