MnemoverseDocs

Appearance

Persistent memory for AI agents ​

Mnemoverse is a persistent memory API for AI agents: statistical learning, not vector search.

Sign up free at console.mnemoverse.com — no credit card, 1,000 queries/day.

3 Lines of Python ​

python
from mnemoverse import MnemoClient

client = MnemoClient(api_key="mk_...")

# Store a memory
client.write("Retry with exponential backoff fixed the timeout issue",
             concepts=["retry", "backoff", "timeout"])

# Query — Hebbian associations expand "timeout" → "backoff", "retry"
results = client.read("how to handle timeouts?")

# Report outcome — system learns what works
client.feedback(atom_ids=[r.atom_id for r in results.items], outcome=1.0)

Not a Vector Database ​

MnemoversePinecone / Weaviate / Chroma
Core modelHebbian associations + prediction-error feedback (Rescorla-Wagner-style)Vector embeddings (cosine similarity)
Learns from outcomesYes — feedback loop updates valenceNo — static retrieval
Concept associationsThree-factor Hebbian graphNone
Memory compressionAvailable in the engine, deliberately off on the hosted serviceRetention is application-managed (TTL or explicit deletes)
Query expansionAutomatic via learned associationsManual or none
Starting priceFree (1,000 queries/day)$25-70/mo

We don't just store and retrieve. We learn which memories matter.

How It Works ​

text
1. WRITE  →  Novelty gate refuses a write too close to one already stored  →  Stored with semantic embedding
2. READ   →  Hebbian expansion  →  Valence-boosted ranking  →  Recency boost  →  Results
3. FEEDBACK  →  Outcome updates valence  →  Strengthens associations

Consolidation, which groups similar memories under a prototype, is designed but switched off on the hosted service; see the API reference.

The association graph itself is readable, not just used internally to expand a query: POST /memory/graph returns the Hebbian edges — weight, valence, count — around a set of seed concepts, from the REST API, the Python SDK's graph(), or the MCP tool memory_graph (package 0.13.0 and the hosted connector).

Memories that lead to good outcomes rank higher in future queries. Memories that consistently fail rank lower; feedback changes the order of the next read, not the stored text. The system improves with use.

Pricing ​

FreeProTeam
Price$0/mo$29/mo$149/mo
Queries/day1,00050,000500,000
Atoms (memories)10,000500,0005,000,000
Rate limit60/min600/min3,000/min
API keys1UnlimitedUnlimited
SupportCommunityPriorityPriority

Built on Research ​

  • Published: SLoD framework on arXiv (Semantic Level of Detail)
  • Benchmarked: evaluated on LoCoMo (1,986 questions) and LongMemEval — see Benchmarks for the current numbers
  • Validated: experiments across game-playing agents and academic benchmarks

Common questions ​

What is persistent memory for AI agents? ​

Persistent memory lets an AI agent store facts, preferences, and decisions once and recall them in later sessions, instead of starting cold every time. Mnemoverse is a hosted persistent-memory API: you write a memory through one key or OAuth sign-in, and any connected tool — Claude, Cursor, VS Code, ChatGPT, or a Python script — can read it back later.

How do AI agents get persistent memory? ​

Connect Mnemoverse and the agent gains 11 tools: the core memory tools, Beta shared-room tools and a vault alias listing; the full list is on the MCP server page. In ChatGPT the connector offers 10, without vault_list (where each tool works). The agent calls memory_write to save a durable fact and memory_read to recall it on a later turn or in a new session. You get an API key (mk_live_...) free at console.mnemoverse.com — no credit card, 1,000 queries/day. See Getting Started for the first call, and Make Your Agent Use Memory for getting the agent to actually call the tools. Memory can also be shared across separate accounts through a room (Beta), instead of staying private to a single key.

A vector database does static similarity retrieval: you store embeddings and get back the nearest matches, the same way every time. Mnemoverse adds a learning layer on top of retrieval — Hebbian associations expand a query to related concepts, and a prediction-error feedback signal (Rescorla-Wagner-style) updates each memory's valence so memories that led to good outcomes rank higher next time. Retrieval is the floor, not the whole system. See Not a Vector Database above for the side-by-side.

Does the memory learn over time? ​

Yes. After a recalled memory helps or misleads, the agent calls memory_feedback; that outcome updates the memory's valence, so useful memories surface faster and consistently unhelpful ones rank lower; feedback changes the order of the next read, not the stored text. Memories are kept rather than compressed away: consolidation exists in the engine but is switched off on the hosted service, because it has never once completed in production and a prototype competes with its own sources for the same result slots.

Next Steps ​