MemoryAtlas

Encodes conversation turns as hierarchically structured facts — world facts and experience facts consolidate into observations that merge into mental models, with automatic deduplication and evidence-tracking across updates. TEMPR retrieval runs four strategies in parallel (semantic, keyword/BM25, entity-graph, temporal) and merges results before returning them to the agent. Reflect reasons over retained memories using configurable mission, directives, and disposition settings per memory bank. Iris Extract provides a separate LLM-mediated structured-extraction pass for document ingestion. Memory banks are isolated namespaces; 40+ official framework adapters plus a first-party MCP server expose all operations.

Storage
Embedded Postgres backing four search indexes (semantic vector, BM25 keyword, entity-relationship graph, temporal store); deployed via Docker Compose, Helm chart, or pip/uvx. Hindsight Cloud provides a fully managed tier with daily backups, analytics dashboard, and 99.9% uptime SLA.
Retrieval
TEMPR: four strategies run in parallel — semantic similarity, keyword/BM25 for exact names and technical terms, entity-graph traversal for relationship queries, and temporal matching for time-anchored queries (e.g. 'last spring'); results merged and ranked before delivery. Mental Model Retrieve is a separate lower-cost path for high-level context.
Self-host
Self-host: moderate
License
MIT
Pricing
Self-host free, MIT licensed, no limits (Docker Compose / Helm / pip). Hindsight Cloud pay-as-you-go (repriced as of 2026-07-23): Retain $10.00/M tokens (was $15/M), Recall $0.75/M, Reflect $0.05/call (was $3/M — now billed per call, not per token), Iris Extract $7.50/M, Mental Model Retrieve $0.25/M, Mental Model Refresh $0.05/call (was $3/M — now per call), plus a new Storage line at $0.25/M tokens/month for memories held past their first 30 days (first 30 days free); free starter credits, no fixed monthly fee. Enterprise: custom pricing, dedicated infrastructure, up to 24x7 SLA, SSO and RBAC included. · Free + paid
GitHub stars
19,457
Last release
2026-08-07
Last commit
2026-08-10
First catalogued
2026-06-28

Strengths

  • Strongest neutral benchmark scores on LongMemEval among catalogued frameworks
  • TEMPR parallel retrieval combines semantic, BM25, entity-graph, and temporal search simultaneously for broad recall
  • Reflect synthesis configurable via per-bank mission/directives/disposition — gives each agent a distinct memory personality
  • 40+ official framework integrations (LangGraph, CrewAI, LlamaIndex, OpenAI Agents SDK, Zapier, Continue.dev, n8n, and more)
  • Memory Defense security layer; OAuth 2.1 on Hindsight Cloud; audit logs, webhooks, and RBAC for enterprise use

Watch out

  • Heavier self-host than append-only layers: four indexes (vector + BM25 + graph + temporal) running inside an embedded Postgres
  • Cloud retain cost ($15/M tokens) can add up quickly at high write volume — benchmark your use case against the free self-hosted path first
  • Remote MCP transport (remote-http via Hindsight Cloud) requires a cloud account; local deployment uses HTTP over localhost, not stdio

Best for

  • Apps where recall accuracy is the priority (strong LongMemEval scores)
  • Persistent memory for coding agents (Claude Code / Cursor / Continue.dev / OpenCode / Roo Code)
  • Teams needing enterprise-grade memory: bank isolation, audit logs, webhooks, SSO, RBAC, and a managed cloud option

How it integrates

Benchmark results

BenchmarkValueBackboneTrustSource
beam-10m64.1 accuracySelf-reportedHindsight (Vectorize)
longmemeval91.4 accuracyGemini 3 ProSelf-reportedHindsight (Vectorize)

Sources

Last verified 2026-08-01 · updated by refresh-framework-cards