Skip to content

Self-Improving AI Memory

AI Memory That Gets Smarter Over Time

The more you use MetaMemory, the better it retrieves. Self-improving retrieval powered by Thompson Sampling, meta-learning, and online adaptation. Works across Claude, ChatGPT, Gemini, and Cursor.

Documentation
84.4%
LongMemEval-S
Overall, 500 questions
89.0%
Task-Averaged
vs GPT-4o 60.6%
0.75→1.0
Recall@10
Learning on vs static
+0.24
MRR Lift
Channel weight learner
How It WorksEncode · Retrieve · Learn
1. Encode

Your agent stores a memory. MetaMemory encodes it with a semantic embedding plus emotion tags, temporal metadata, keywords, and graph entities.

2. Retrieve

On search, 5 channels run in parallel — semantic, temporal, emotional, keyword, graph — and results are fused via Reciprocal Rank Fusion.

3. Learn

Thompson Sampling + meta-learning + online adaptation pick the best strategy per context. Quality improves with every query you make.

Self-improving retrieval. The more you use it, the smarter it gets.

Why Memory Matters

AI agents need more than context windows

Memory That Degrades

Append-only vector stores get noisier over time. Retrieval quality declines as bad memories pile up.

MetaMemory gets smarter as you use it. Thompson Sampling learns which retrieval strategy works best per context. Meta-learning rules accumulate from your queries. Online learning adapts in real time.

RAG Isn't Memory

RAG retrieval is static — it never learns from what worked. Same query, same ranking, forever.

5-channel fused retrieval (semantic + temporal + emotional + keyword + graph) with adaptive weighting per query type. The system updates its own strategy based on every search you make.

Agents Forget

Traditional agents lose context between sessions. Every conversation starts from zero — even across platforms.

Episodic storage with 3-level isolation (agent, session, user). Memories persist across sessions, restarts, deployments — and across Claude, ChatGPT, Gemini, Cursor with one account.

Facts Go Stale

Your user changed jobs. The vendor got swapped. The decision was reversed. But retrieval keeps serving last month’s truth next to this week’s.

A bi-temporal claims ledger extracts facts from memories and tracks when they were true and when the system learned them. Updates supersede instead of contradict — exactly one live current fact per claim, with full point-in-time history.

Features

Self-improving memory, end to end

Three things set MetaMemory apart: retrieval that improves with use, facts that stay current, and shared memory that agents can't corrupt. Everything else is what makes them work.

Adaptive Strategy Selection

Thompson Sampling and UCB algorithms learn which retrieval strategy works best for each query type — so recall climbs with use instead of degrading.
Explore Adaptive Strategy Selection

Bi-Temporal Claims Ledger

Facts are extracted as claims and tracked on two time axes, so knowledge updates supersede cleanly and you can query what was known at any point in time.
Explore Bi-Temporal Claims Ledger

Multi-Agent Arbitration

When long-running agents disagree about a fact, a trust-weighted arbitration ladder decides which version stands — no silent overwrites, every decision audited.
Explore Multi-Agent Arbitration

How It Works

Layered architecture, simple interface

How It Works

Three stages. One seamless pipeline.

01

Encode

Every interaction is encoded with a semantic embedding plus emotion tags, temporal metadata, keywords, and graph entities. This captures not just what was said, but its meaning, context, and feeling.

  • Semantic embeddings via your chosen provider
  • Emotional state detection and tagging
  • Automatic episode boundary detection
02

Consolidate

LLM-powered consolidation merges related memories, compresses redundant information, and strengthens important connections — just like sleep does for the human brain.

  • Fact extraction into a bi-temporal claims ledger
  • Redundancy compression with semantic preservation
  • Cross-session memory linking
03

Retrieve

5 specialized retrieval channels compete and collaborate using Thompson Sampling to surface the most relevant memories for each query, learning and improving over time.

  • Adaptive strategy selection via multi-armed bandits
  • Gradient-boosted relevance ranking
  • Recall that improves with every query

BYOK Providers

Bring your own keys

Use your existing API keys from any supported embedding provider. Keys are validated, encrypted at rest, and rotatable at any time — you control the provider, the model, and the spend.

OpenAIEmbeddings + LLM
Google GeminiEmbeddings + LLM
CohereEmbeddings
Voyage AIEmbeddings
Azure OpenAIEmbeddings + LLM
MistralEmbeddings + LLM
OllamaLocal
QwenSelf-hosted

Cognitive Architecture

Built on research, not heuristics

50

Retrieval Channels

Semantic, temporal, emotional, keyword, and graph — fused via Reciprocal Rank Fusion

20

Time Axes Per Fact

Bi-temporal claims: when it was true, and when we learned it

5000

LongMemEval Questions

Full LongMemEval-S evaluation: 84.4% overall accuracy

60

Emotion Categories

Confident, uncertain, confused, frustrated, insight, breakthrough

Memory that gets smarter over time

Two lines of MCP config. Bring your own provider keys — validated, encrypted, under your control. Your AI remembers across sessions — and the system learns which retrieval works best for you.