The Mnemosyne approach
Find the evidence. Know when it held. Test what deserves reuse. Mnemosyne turns these requirements into one inspectable lifecycle, from your first source to the next agent’s decision.
Mnemosyne Recall
Retrieval quality matters twice: when memories are selected, and when they must fit into the next turn.
Default local scoring accounts for rare terms, repetition and document length using only authorized, time-eligible records under the query’s trust settings. Optional semantic retrieval searches the full indexed scope, then fusion and optional reranking select candidates under explicit limits.
Mnemosyne On-Device Intelligence runs embeddings and reranking on your CPU. Provision runtime and model weights explicitly, then use their cache offline. No model is needed for the local store or default lexical search.
Explore on-device intelligenceExpand an issued source handle into the exact original text in bounded pages. Corrections and changed evidence invalidate affected generated context and old handles.
Explore adaptive contextPersistent facts and clear ownership
Preserve useful context with its source, scope and uncertainty intact.
Capture selected documents and visible conversations with stable source identities. Keep records private by default and share deliberately across agent identities in the same workspace.
Typed profiles keep supported values and competing candidates tied to their sources. Preview additions and corrections in private change sets, then merge against current evidence.
Explore evidence-backed profilesMnemosyne Timeline
Separate when a fact held in the world from when the system learned it. Carry that distinction through recall and compiled context.
An evidenced correction preserves relevant earlier guidance until its effective boundary, then withholds obsolete descendants. Recorded correction causes support historical reads; ambiguous legacy invalidations remain withheld. This is conservative source-level retirement, not yet selective repair of every independent claim.
Explicit source checks distinguish confirmed, changed and unavailable evidence. Action read sets revalidate declared local dependencies immediately before your application considers the action.
Explore freshness and action checksMnemosyne Evidence Gate
Turn selected experience into reusable guidance with a visible path from proposal to trial to outcome.
Candidate skills accumulate controller-run trials before becoming ordinary advice. The recommended policy requires at least two distinct tasks and two distinct verifiers; the compatibility default remains one of each. Saved requirements survive restarts and cannot be lowered by weaker reopening options.
Attribute success and failure to the memories actually used. Recheck supporting evidence before reuse, and retire dependent guidance when its sources change or fail.
Explore learning and validationMnemosyne Migration
Start beside the memory system your agents already use, and increase adoption as evidence accumulates.
Connect a read-only adapter to existing search. Copy records your agents retrieve, then use policy-gated paired recall checks to move from shadow to assist to local preference. The old store remains checked on every query and is never changed by the bridge.
Return routing to the existing system when needed. Choose a reviewed full export import, inspect source bytes, undo unchanged batches and rehearse whole-database recovery. Progress reports cover the queries observed so far.
Explore gradual migration and rollbackEvidence before claims
A frozen, matched 500-question corpus measures our recall upgrade against our earlier scorer. It does not establish a ranking against other products.
All annotated source sessions were retrieved for 425 of 470 answerable questions, up from 386 (82.13%) with the earlier scorer. Both conditions use the same corpus and top-20 chunk limit; all 500 questions were processed.
Under the same 8,192-byte context budget, complete source-session coverage rises from 367 to 382 of 470 answerable questions. Stronger scoring adds CPU work; the protocol reports retrieval timing alongside coverage.
These are session-level evidence scores, not generated-answer accuracy. A returned chunk can earn source-session credit without containing the answer sentence. Same-day-compatible timing permits later sources on the question’s calendar day, so this is not a strict no-future-evidence experiment.
Inspect the protocol and limitsThe separate agent harness compares no memory, recent history, lexical memory and adaptive context using a matched reader and budgets. It can track answer quality, grounding, stale actions and negative transfer. Scripted examples check the mechanics; generated-answer quality still needs measurement with your reader and tasks.
Explore the evaluation protocolFollow a source through capture, reuse, correction and the next agent handoff.