mnemosyne-oss/mnemosyne

Selection

Select when

• You need persistent cross-session memory in Claude Code through MCP.
• You prefer local SQLite storage over a hosted memory platform.
• You want one memory layer shared by Claude Code and other MCP or Python agents.
• You need explicit remember, recall, forget, export, and import operations.
• You want hybrid FTS5 and optional vector retrieval without an external vector database.
• You need scoped memories, expiry, banks, or temporal triples.
• You are willing to operate a self-hosted sync endpoint for cross-device memory.
• You require optional client-side encryption for sync payload content.
• You want a deeper first-party memory-provider integration with Hermes Agent.

Boundaries

Avoid when

• You require a native Claude Code plugin with lifecycle hooks or automatic capture evidenced for Claude Code.
• You cannot allow an agent-accessible store to retain sensitive conversation or project content.
• You need independently audited encryption, authorization, sandboxing, or privacy guarantees.
• You require independently reproducible state-of-the-art memory benchmarks.
• You need zero operational dependencies while also requiring vector search, MCP, local LLMs, or sync.
• You cannot review the privacy effects of remote embedding or LLM endpoints.
• You need a managed multi-tenant service with documented enterprise controls.
• You require proven high-quality multilingual retrieval without model selection and evaluation.
• You need guaranteed sub-millisecond retrieval at large corpus sizes.
• You cannot manage backups and encryption keys for local or synchronized SQLite data.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only selected first-party files were supplied, so MCP schemas, retrieval code, encryption implementation, authorization, migrations, tests, packaging, and Claude Code integration behavior were not directly inspected.
• Uncertainty: repository and benchmark dates extend into 2026 relative to this analysis context, so recency, release status, adoption counts, and version claims cannot be independently reconciled.
• Uncertainty: the dedicated Claude Code integration guide was referenced but not supplied, leaving config-file location, minimum host version, and tested compatibility unverified.
• The files report different tool counts across basic MCP, full MCP, legacy Hermes, and current Hermes surfaces, so exact availability is version and integration dependent.
• The README withdraws a prior LongMemEval result for lack of methodology, while its comparison table still lists a different LongMem figure, creating unresolved benchmark-documentation inconsistency.
• The documentation calls the system sub-millisecond, but the supplied pure-retrieval table reports 35–493 ms, so latency claims require workload-specific validation.
• The published BEAM result is explicitly point-in-time, has not been rerun on the current tree, and uses a different judge from a compared competitor.
• The reported pure-retrieval Recall@10 is 20%, which the project itself describes as low.
• Remote embeddings and LLM-driven extraction can transmit stored memories and recall queries to third-party endpoints.
• Local SQLite files can contain sensitive preferences, credentials, conversations, and project details; local at-rest protection and host access controls require careful configuration.
• Sync without client-side encryption exposes memory content according to the operator's infrastructure, while loss of an encryption key can make encrypted memories unrecoverable.
• No independent security audit, privacy assessment, Claude Code compatibility matrix, benchmark reproduction, or production reliability study was supplied.

The default documented Claude Code path launches a local stdio MCP process over a local SQLite store. Network exposure appears optional through SSE or Streamable HTTP MCP, remote embeddings or LLMs, and sync. Non-loopback HTTP requires allowed-host and, for browsers, allowed-origin configuration according to the README. Sync supports API keys, JWTs, TLS guidance, and optional authenticated client-side payload encryption. The evidence does not establish audited implementation quality, local database encryption defaults, per-tool authorization, or safe handling of agent-supplied sensitive content.