Carry knowledge forward
Capture information and retrieve it in later sessions. Start locally; choose a deployment backend for your workload.
EXTERNAL MEMORY FOR AI AGENTS
Memory with a
source of truth.
Give an AI agent information it can carry between sessions—and a way to trace, inspect and revise what it remembers.
Local-first · Evidence-led · In active development
A CONTINUING RECORD
01 Capture evidence↓02 Build useful projections↓03 Retrieve with provenance
The source stays distinct from its interpretation.
BEYOND A SINGLE CONVERSATION
An agent’s context ends.
Its memory should remain useful.
Mnemosyne is an external AI memory system. It stores information outside the model’s weights and active context, with mechanisms to reuse and manage it across interactions. A larger context window alone does not provide that lifecycle.
Keep the evidence
Content-addressed evidence and provenance separate the original record from derived beliefs and interpretations.
Make revision explicit
Time-aware beliefs, supersession and retraction primitives represent changing information instead of treating every remembered statement as permanently true.
DESIGNED TO BE INSPECTED
Evidence in. Useful memory out.
Capture → project → retrieve
The evidence ledger is the foundation. Derived projections organize information for retrieval and reasoning. Search results can carry provenance and confidence so applications can decide how to use them.
A choice of interfaces and backends
CLI and Model Context Protocol interfaces connect memory to applications. Local, SQLite and PostgreSQL backends serve different deployment needs; advanced retrieval depends on the selected backend and configuration.
Control is part of memory
Tenant boundaries, capability-mediated writes, branchable memory and deletion primitives are part of the design. Each surface needs its own validation; these mechanisms are not a blanket production-security certification.
A PRODUCT AND ITS PROOF ARE DISTINCT
Meet Mnemetric.
The benchmark platform evaluates external memory systems—including Mnemosyne—with visible protocols, source mappings and limitations. Its aim is broad lifecycle coverage, not a predetermined winner.
Mnemetric is operator-run. Mnemosyne is an affiliated entry, not an independently certified leader.
WHERE WE ARE TODAY
Built in the open. Measured with care.
Usable core
Core local capture and retrieval are usable. Advanced capabilities and deployment profiles remain under development and require configuration-specific validation.
Results have boundaries
Measurements apply to the recorded source version, dataset, protocol and hardware. Development checks are not full official benchmark results or proof of superiority.
Important open work
Windows validation has gaps. Complete source lineage and provider or parametric unlearning remain open work. Deleting local evidence does not establish deletion from backups or model weights.