Mnemosyne — Native AI Memory for Hermes Agent | Zero-Dependency, Sub-Millisecond<br>SupportWe need your help.Read our story →

English<br>DiscordGitHubpip install

v3.0.0 — MEMORIA Fact Engine<br>Memory for Hermes<br>The native memory system for Hermes Agent. SQLite-backed, sub-millisecond, zero dependencies. No cloud. No API keys. Just pure speed.<br>Get StartedRead the Docs<br>Migrate from Zep, Mem0, Honcho, or Hindsight in one command — see migration docs<br>Query latency<br>0Dependencies<br>98.9%LongMemEval<br>100%Local & private

SimpleThree lines. Infinite memory.<br>No configuration files. No environment variables. No cloud accounts. Import, remember, recall. That is all.<br>pip install mnemosyne-memory<br>Zero external services required<br>Works offline, always<br>Hermes Agent integration built-in

from mnemosyne import remember, recall

# Store a memory<br>remember(<br>"User prefers dark mode",<br>importance=0.9,<br>scope="global"

# Retrieve relevant context<br>results = recall("user preferences")<br># => [{"content": "User prefers dark mode", ...}]

FeaturesEverything you need. Nothing you do not.<br>Built from the ground up for AI agents that need fast, reliable, persistent memory.

Sub-Millisecond Latency<br>Direct SQLite access delivers &lt;1ms queries. No network overhead. No HTTP roundtrips.

100% Private<br>All data stays on your machine. No cloud services. No data leaves your device, ever.

Native Vector Search<br>sqlite-vec integration for semantic search. Hybrid ranking: 50% vector + 30% FTS + 20% importance.

Beam Architecture<br>Three-tier memory: working_memory for hot context, episodic_memory for long-term, scratchpad for reasoning.

Auto Consolidation<br>Old working memories are automatically summarized and moved to episodic storage via sleep cycles. Configurable auto_sleep intervals.

Hybrid Search<br>Combines vector similarity, full-text search, and importance scoring for the best recall accuracy.

Streaming & DeltaSync<br>Real-time incremental memory updates via DeltaSync. Stream results as they arrive — no more waiting for full batches.

Smart Filtering<br>ignore_patterns blocks noisy or irrelevant content from entering memory. Keep your context window clean and focused.

SpeedNumbers that speak<br>Measured on CPU with sqlite-vec + FTS5. No GPU required.

Write0.81ms56x faster<br>Read0.076ms500x faster<br>Search1.2ms43x faster<br>Cold Start0msInstant

OperationHonchoZepMem0MnemosyneWrite45ms85ms50ms0.81msRead38ms62ms45ms0.076msSearch52ms78ms60ms1.2msCold Start500ms800ms300ms0ms<br>BEAM Benchmark (ICLR 2026)<br>End-to-end memory retrieval at scale. LLM-as-judge against published baselines.

100K Context35.4%Retrieval from 100K-token conversations<br>500K Context19.3%Retrieval from 500K-token conversations<br>1M Context19.2%Retrieval from 1M-token conversations

CompareMnemosyne vs. cloud memory providers<br>See exactly what you gain — and what you trade — when you switch.

FeatureMnemosyneHonchoZepMem0CostFree forever$$$ Paid (credits)$$$ Paid (Flex+)Freemium ($0-$249/mo)HostingLocal - your machineCloud onlyCloud / BYOCCloud onlyPrivacy100% local, zero exfilExternal API callsExternal API callsExternal API callsOffline modeYes - airplane modeNoNoNoSetuppip installDocker + API keysDocker + PostgresAPI key + signupVector storesqlite-vec (built-in)pgvector (external)pgvector (external)pgvector (external)Full-text searchFTS5 (built-in)Separate serviceSeparate serviceSeparate serviceAuth requiredNoneSupabase authOAuth / API keyAPI keyRate limitsUnlimitedPlan-dependentCredit-basedPlan-dependentData ownershipYou own the SQLite fileVendor-hostedVendor-hostedVendor-hostedExport / importOne JSON fileLimitedLimitedLimitedDependenciesPython stdlib + ONNXDocker, PostgresDocker, Postgrespip + API keyMemory architectureBEAM (3-tier)Session + factsGraph RAG + factsSession + factsAuto-consolidationSleep cycles built-inManual / paidManualManualTemporal triplesNative with validityNoNoNoLongMemEval98.9% Recall@All@5Not publishedNot publishedNot publishedBEAM-100K35.4% / 19.3% / 19.2%Not publishedNot publishedNot published<br>Switching from Honcho<br>You gain500x faster reads, zero monthly bill, 100% offline, no Docker, no credit system

You loseCloud dashboard, managed scaling, team sharing

Switching from Zep<br>You gain43x faster search, no PostgreSQL to maintain, no deployment overhead, instant cold start

You loseGraph RAG viz, SOC 2 certs, managed BYOC

Switching from Mem0<br>You gainSub-millisecond everything, no rate limits, no vendor lock-in, full data portability

You loseManaged platform, 90K+ community, YC ecosystem

Switching from Hindsight<br>You gainZero dependency, no network calls, SQLite-native, BEAM architecture

You loseCloud sync, managed inference, web dashboard

The bottom line<br>✓Speed: 43-500x faster than cloud alternatives — zero HTTP roundtrips.<br>✓Privacy: Data never leaves your machine. No API calls. No telemetry.<br>✓Cost: Zero ongoing cost. No credits. No tiers. No "contact sales."<br>✓Simplicity: One pip install. No Docker. No config. No signup.

Trade-off: You manage your own backup (one SQLite file). No web...