Mnemosyne is live · free · last verified 25 Oct 2026 · viability healthy
Mnemosyne
Local-first AI agent memory in one pip install: SQLite-backed recall with sub-millisecond queries and no cloud account.
For single-machine agents this is the memory layer to try first: sub-millisecond local queries, zero network hops, one SQLite file you actually own, MIT-licensed with no recurring cost. The v3.0.0 BEAM results (65.2% end-to-end at 100K, 20% Recall@10 holding to 10M, 100% abstention) edge out Honcho's 63.0% on the paper baseline. What you give up is real: no cloud sync, no dashboards, no SLA, and backups are your problem. If you need Zep's managed graph or a team console, stay put.
Verdict generatedVerified 5d ago · liveness 70/100 · cite: rightaichoice.com/tools/mnemosyne
- AI agent developers who want fast, private memory on one machine
- Teams with strict privacy or offline/air-gapped requirements
- Hermes, Claude Code, Cursor or Codex users adding persistent recall
- Researchers and hobbyists prototyping memory-augmented agents on a zero budget
- Developers needing cloud sync or shared memory across multiple machines
- Teams wanting managed hosting, dashboards or auto-scaling
- Organizations requiring SOC 2 certification or vendor SLAs
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Mnemosyne if you need shared memory across several machines, a hosted dashboard, user auth and access control, or a vendor SLA behind your agent's recall — it is a single-machine SQLite library by design.
You own the backup: if the SQLite file is lost or corrupted, there is no vendor-side restore, so plan for your own snapshot routine.
Mnemosyne is MIT-licensed with no tiers, credits or usage meter, so the comparison is against recurring cloud memory bills rather than against a higher tier of Mnemosyne. Honcho and Zep sit in paid-credit and paid-Flex territory on the vendor's own comparison table, and Mem0 is listed there as freemium ($0-$249/mo). For a solo developer on one machine, Mnemosyne's real cost is your time on backups.
In short
Mnemosyne — Local-first AI agent memory in one pip install: SQLite-backed recall with sub-millisecond queries and no cloud account. Best for AI agent developers who want fast, private memory on one machine, Teams with strict privacy or offline/air-gapped requirements, Hermes, Claude Code, Cursor or Codex users adding persistent recall. Free to use.
What is Mnemosyne?
Mnemosyne — Local-first AI agent memory in one pip install: SQLite-backed recall with sub-millisecond queries and no cloud account. Best for AI agent developers who want fast, private memory on one machine, Teams with strict privacy or offline/air-gapped requirements, Hermes, Claude Code, Cursor or Codex users adding persistent recall. Free to use.
Mnemosyne is a persistent memory layer for AI agents that runs entirely on your own machine. A single `pip install mnemosyne-memory` puts a pure-Python memory layer on disk, backed by SQLite — no Docker, no Postgres, no API keys, no cloud account, no credit system. The whole API is two calls: `from mnemosyne import remember, recall`. Underneath sits BEAM (Bilevel Episodic-Associative Memory), three SQLite tables doing different jobs. working_memory holds hot context that gets auto-injected into prompts, episodic_memory stores long-term memories searched with sqlite-vec and FTS5, and scratchpad carries temporary agent reasoning that clears per session. Recall uses hybrid ranking weighted 50% vector similarity, 30% full-text and 20% importance, so a fuzzy semantic match and an exact keyword both surface. Configurable sleep cycles summarize aging working memories into episodic storage on their own, and ignore_patterns keeps junk out of your context window. Published numbers for v3.0.0 (May 2026, against ICLR 2026 baselines): 65.2% end-to-end QA at 100K context under LLM-as-judge, Recall@10 holding at 20% as the corpus grows to 10M, and 100% abstention when the corpus holds no answer. On CPU, write is 12ms at 1K entries and FTS5 search 5ms. The project is MIT-licensed, has 3k+ GitHub stars and is funded through an OpenSource Collective on OpenCollective. It's for developers wiring memory into a single-machine agent who would rather own the SQLite file than rent a vector database. The trade-off is explicit: you handle your own backups. Compared with Honcho (63.0% on the BEAM paper), Zep and Mem0, Mnemosyne replaces Docker-plus-pgvector with a file — and gives up cloud dashboards, team sharing and vendor SLAs to do it.
Key features
- Sub-millisecond query latency via direct SQLite access, no HTTP roundtrips
- 100% local and private — no telemetry, no data leaves your device
- Native vector search through built-in sqlite-vec integration
- Full-text search with SQLite FTS5 (5ms at 1K entries)
- Hybrid ranking: 50% vector, 30% full-text, 20% importance
- BEAM three-tier memory: working_memory, episodic_memory, scratchpad
- Automatic consolidation of old memories via configurable auto_sleep cycles
- Streaming results and DeltaSync for real-time incremental updates
- Smart filtering with ignore_patterns to block noisy content
- Offline support — works in airplane mode with zero external services
- MCP (Model Context Protocol) support for agent hosts
- Python SDK with remember / recall / beam.sleep API
- Temporal triples with native validity tracking
- Export and import memory as a single JSON file
- Interactive browser demo (client-side simulation, zero LLM calls)
Behind the verdict
Mnemosyne's pitch is narrow and honest, which is refreshing in a category full of managed memory platforms. The architecture is three SQLite tables with different jobs: working_memory for hot context auto-injected into prompts, episodic_memory for long-term storage searched by both sqlite-vec and FTS5, and scratchpad for temporary reasoning that clears each session. Recall blends 50% vector similarity, 30% full-text and 20% importance, so an exact keyword hit isn't drowned by a fuzzy semantic near-miss. Configurable auto_sleep cycles summarize aging working memories into episodic storage without you writing a job queue, and ignore_patterns keeps noisy content out of the context window. Strengths are speed, privacy and cost. Because there is no server, there is no HTTP roundtrip: the vendor measures write at 12ms and FTS5 search at 5ms at 1K entries on CPU, no GPU required. Nothing leaves your machine — no API calls, no telemetry — which makes it viable for air-gapped and privacy-sensitive work. Export and import is one JSON file, so leaving is cheap. The API surface is genuinely two functions, `remember` and `recall`, with `beam.sleep` for consolidation, and it plugs into Hermes, Claude Code, Cursor, Codex, OpenWebUI and OpenClaw plus MCP for other hosts. Weaknesses are structural, not bugs. It's local by design, so there is no cross-machine sync, no shared memory for a team, no hosted dashboard, and no SLA behind it. You own the backup of the SQLite file. The vendor's only public scale benchmark is v3.0.0 (May 2026): 65.2% end-to-end QA at 100K context under LLM-as-judge, 20% Recall@10 out to 10M corpus, and 100% abstention when the answer isn't in the corpus. That's competitive with Honcho's 63.0% on the BEAM paper, but it's one benchmark, and the interactive browser demo on the site is an explicit client-side simulation with zero LLM calls — it shows the shape of the workflow, not production recall quality. Where it fits: a developer running a single agent on one machine who wants persistent recall without renting a vector database. Where it doesn't: teams needing shared memory, user auth and access control, SOC 2 paperwork or a vendor on the phone when something breaks — for those, Zep or Mem0 remain the safer hire despite the recurring bill.
Limitations
- Mnemosyne is a pure-Python, SQLite-backed memory layer that runs entirely locally, so it ships no cloud service and depends on a host agent framework (Hermes, Claude Code, Cursor, Codex, OpenWebUI, OpenClaw) plus MCP for integration.
- The vendor's only published benchmark at scale is for v3.0.0, reporting 65.2% end-to-end QA at 100K context and 20% Recall@10 holding to 10M corpus scale.
- Because all data stays on the local machine, backup and multi-device concerns are your own, and the interactive browser demo is an explicit client-side simulation with zero LLM calls rather than the real product.
- There is no hosted team collaboration: Mnemosyne is built for individual agents and single-machine deployments.
as of 2026-10-20
Researching Mnemosyne? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Is Mnemosyne still active?
Mnemosyne is live · free · last verified 25 Oct 2026 · viability healthy.
Viability score
How well maintained and how widely used is Mnemosyne? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →What do users say about Mnemosyne?
What people actually say about Mnemosyne — is it worth it?
Our community sentiment research written Jul 31, 2026 is more than 90 days old, so it is being re-reviewed against current data and is not shown.
How we verified this
Every listed fact about Mnemosyne— pricing, features, links, status — is re-read from the vendor's own pages by our verification pipeline and each change is logged with a date; the last pass ran on 25 Oct 2026. Pages mix model-assisted synthesis with human review; see how we verify.
Verification history
We have re-verified Mnemosyne 7 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — changed SSL certificate, summary, hidden costs, pricing fit, our verdict, who should skip it, features, switching out, switching in, limitations, description, setup time
- — changed summary, our verdict, description, pricing tiers, who should skip it, features, who it suits
- — changed limitations
- — changed hidden costs, limitations, switching in, features, switching out, who should skip it, our verdict, pricing tiers, pricing fit, setup time, description, who it suits, summary, use cases
- — changed pricing page, limitations
- — changed summary, who it suits, who should skip it, pricing tiers, features, description, our verdict
Showing the 6 most recent of 7 verification passes.
Free to cite with attribution — this page re-verifies continuously.
How much does Mnemosyne cost in 2026?
Mnemosyne: Free to use. Pricing last verified 25 Oct 2026.
Plans compared
Every published Mnemosyne tier, its price as the vendor publishes it, who it fits, and what it adds over the tier below. The 12-month cost projection follows.
| Plan | Price | Billing | Who it's for |
|---|---|---|---|
| Free | Free | — | Solo developer or researcher wiring persistent memory into one agent on one machine, with no budget and no appetite for a cloud account.Adds: Starting tier: MIT-licensed and free forever, one pip install of mnemosyne-memory, local SQLite with sqlite-vec and FTS5, BEAM three-tier memory, and no API keys, accounts or credit system. |
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Where the pricing makes sense
The company stage and team size where Mnemosyne's pricing actually pencils out — and where peers do it cheaper.
Mnemosyne is MIT-licensed with no tiers, credits or usage meter, so the comparison is against recurring cloud memory bills rather than against a higher tier of Mnemosyne. Honcho and Zep sit in paid-credit and paid-Flex territory on the vendor's own comparison table, and Mem0 is listed there as freemium ($0-$249/mo). For a solo developer on one machine, Mnemosyne's real cost is your time on backups.
Has Mnemosyne's pricing changed?
No plan price for Mnemosyne has moved since we began logging in Aug 2026; the pricing page is re-read on every verification pass.
Who should pick Mnemosyne?
Pick Mnemosyne if you are AI agent developers who want fast, private memory on one machine, Teams with strict privacy or offline/air-gapped requirements, Hermes, Claude Code, Cursor or Codex users adding persistent recall, Researchers and hobbyists prototyping memory-augmented agents on a zero budget.
Look elsewhere if you are Developers needing cloud sync or shared memory across multiple machines, Teams wanting managed hosting, dashboards or auto-scaling, Organizations requiring SOC 2 certification or vendor SLAs, Anyone unwilling to manage their own backups of the SQLite file.
Skip it if Skip Mnemosyne if you need shared memory across several machines, a hosted dashboard, user auth and access control, or a vendor SLA behind your agent's recall — it is a single-machine SQLite library by design.
GeneratedReal-world workflow fit
Concrete scenarios for the personas Mnemosyne actually fits — and what changes day-one when you adopt it.
You pip install mnemosyne-memory, wire MCP into your host, and call remember("User prefers dark mode", importance=0.9, scope="global") on the first turn, then recall("user preferences") on later sessions.
Outcome: Preferences survive between sessions with no account, no API key and no server, and recall returns in under a millisecond because it's a local SQLite read.
You run an agent with no internet access at all and point Mnemosyne at a local SQLite file; ignore_patterns filters log noise out of the store and auto_sleep folds stale working memory into episodic storage each cycle.
Outcome: Persistent recall works in airplane mode with zero external services, no telemetry and no data exfiltration risk.
You need a controllable memory backend to reproduce retrieval behavior, so you use the BEAM three-tier store with hybrid ranking (50% vector, 30% FTS, 20% importance) and native temporal triples with validity tracking, then export everything as one JSON file per run.
Outcome: Each experiment is a self-contained file you can diff and archive, without credits, rate limits or signup friction standing between runs.
Use cases
- Store user preferences persistently during an agent session and recall them later
- Let a conversational agent remember context from previous sessions
- Build a local knowledge graph from agent reasoning without a cloud service
- Consolidate aging context automatically so the prompt window stays focused
- Run an AI assistant fully offline with no internet dependency
- Swap a cloud memory provider (Mem0, Zep, Honcho) for local SQLite recall
- Prototype memory-augmented agents without credits, rate limits or signup
Setup time & first value
How long it actually takes to get something useful out of Mnemosyne — broken out by persona, not the marketing-page minute.
One pip install and roughly three lines of code gets a solo developer to first recall in minutes. Adding MCP wiring to a host like Claude Code or Cursor is a short configuration step. Getting consolidation right — auto_sleep intervals and ignore_patterns — takes a week or two of watching real traffic.
Switching to or from Mnemosyne
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From Honcho: install mnemosyne-memory, replay your stored facts through remember(), and accept the loss of the cloud dashboard, managed scaling and team sharing.
- →From Zep: drop the Postgres deployment and load your facts into episodic_memory; you trade Graph RAG visualization and SOC 2 certs for a single local file.
- →From Mem0: export your memories and re-insert them locally to escape rate limits and vendor lock-in; you give up the managed platform and its community ecosystem.
- →From Hindsight: move to the SQLite-native BEAM store for zero dependency and no network calls, giving up cloud sync and the web dashboard.
- ↗To Honcho: export the SQLite store to JSON and push entries through Honcho's API when you need a hosted dashboard and managed scaling.
- ↗To Zep: rebuild memories in Zep if you need Graph RAG visualization, SOC 2 certification or managed BYOC.
- ↗To Mem0: import your JSON export into Mem0's managed platform when a team console and larger ecosystem matter more than local ownership.
Integrations
Resources & guides
Tutorials & Learning
YouTube returned 6 videos for “Mnemosyne”, and we withheld 6: 6 could not be judged, because “Mnemosyne” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Mnemosyne.
Official links
What are the best Mnemosyne alternatives?
The closest alternatives to Mnemosyne we list are Mem0, PMB and TrueMemory. Ranked by direct product-type match, not category overlap.
| Alternative | Pricing | Starts at | Last verified |
|---|---|---|---|
| Mem0 | Freemium | Pricing not published | — |
| PMB | Free | Free | — |
| TrueMemory | Freemium | Pricing not published | — |
Alternatives to Mnemosyne
View all
Mem0
Mem0 is a memory layer that gives AI agents and apps persistent context across sessions and users
PMB
Local-first memory layer that gives your AI coding agent a SQLite file it can recall from between sessions.
TrueMemory
TrueMemory gives CLI coding agents persistent, local-first memory recall in about 55ms — one file, zero cloud.
Tools that pair well with Mnemosyne
Common stack mates teams adopt alongside Mnemosyne, with the specific reason each pairing earns its keep.
Mem0
Mem0 is a memory layer that gives AI agents and apps persistent context across sessions and users
PMB
Local-first memory layer that gives your AI coding agent a SQLite file it can recall from between sessions.
TrueMemory
TrueMemory gives CLI coding agents persistent, local-first memory recall in about 55ms — one file, zero cloud.
Mnemosyne vs its closest rival
Mnemosyne's closest head-to-head is with Spider Cloud — These aren't competitors, so there's no head-to-head winner — pick by the problem you have. If you need live web pages rendered, unblocked, and streamed back as structured JSON for a RAG pipeline, Spider Cloud is the buy, and budget for proxies and AI extraction as separate line items rather than assuming they're bundled. If your gap is agent memory that stays local, costs nothing, and answers in under a millisecond, Mnemosyne is the obvious free install — just don't expect cloud sync, managed hosting, or SOC 2. Many teams would actually run both: Spider to fetch the web, Mnemosyne to remember it. Verdict generated
Head-to-head comparisons
Mnemosyne vs Spider Cloud
These aren't competitors, so there's no head-to-head winner — pick by the problem you have. If you need live web pages rendered, unblocked, and streamed back as structured JSON for a RAG pipeline, Spider Cloud is the buy, and budget for proxies and AI extraction as separate line items rather than assuming they're bundled. If your gap is agent memory that stays local, costs nothing, and answers in under a millisecond, Mnemosyne is the obvious free install — just don't expect cloud sync, managed hosting, or SOC 2. Many teams would actually run both: Spider to fetch the web, Mnemosyne to remember it.
GeneratedMnemosyne vs Genspark
Choose Genspark if you want an all-in-one AI workspace for research, document creation, and no-code automation, especially if you use Google Workspace, Canva, or Figma. Choose Mnemosyne if you are an AI agent developer needing a blazing-fast, fully local memory layer with zero dependencies and total privacy. They solve completely different problems and are not direct competitors.
GeneratedMnemosyne vs Air Ai
If you work in defense logistics and need AI to compress supply chain timelines, Air is the only choice — but it requires enterprise commitment and a sales conversation. If you build AI agents and need a blazing-fast, private, zero-cost memory layer that works offline, Mnemosyne is unbeatable. They serve completely different worlds; pick based on your domain.
GeneratedMnemosyne vs Presto Voice
These are not competing products and no real buyer chooses between them: Presto Voice replaces a drive-thru headset with managed voice AI for QSR franchise networks, while Mnemosyne is a free SQLite memory layer for someone writing an AI agent on their laptop. Buy Presto if you run multiple drive-thrus and want add-on upselling handled at the speaker post under a managed contract (increasingly via Toast). Use Mnemosyne if you want local, private, zero-cost agent memory you install with pip. Evaluating them side by side only makes sense if you happen to both run restaurants and write agents — otherwise pick by the problem you actually have.
GeneratedMnemosyne vs Temporal Ai
If you need crash-proof orchestration for multi-step agent workflows or microservices, pick Temporal AI — it handles retries, state persistence, and human-in-the-loop out of the box. If you want a lightning-fast, fully private memory layer for your AI agent with zero dependencies and no cloud, Mnemosyne is the clear choice. They solve orthogonal problems: Temporal keeps your workflows alive; Mnemosyne keeps your agent’s memory fast and local.
GeneratedFrequently asked questions
Used Mnemosyne? Help shape our editorial sentiment research.