Memory & State

Decide what survives the turn

What should the system still know tomorrow?

13 entries · 2 categories · 3 equations

The arithmetic

Bi-temporal fact

f \;=\; \big(s,\, p,\, o,\; [\,t_{\text{valid}},\, t_{\text{invalid}}\,),\; t_{\text{ingest}}\big)

Two independent time axes: when the fact was true in the world, and when the system learned it. Drop the second and you cannot reconstruct what the agent believed at the moment it acted — which is exactly what an incident review asks for. This interval is Graphiti's actual differentiator over flat fact stores.

Recency-weighted recall

\mathrm{score}(m) \;=\; \lambda\,\mathrm{sim}(q, m) \;+\; (1-\lambda)\,e^{-\Delta t / \tau}

Almost every memory layer implements some version of this. The failure mode is structural: a stale fact with high similarity outranks a fresh correction, because exponential decay only discounts age — it never marks the old fact false. Temporal invalidation does.

Memory selection is a knapsack

\max \sum_i v_i x_i \quad \text{s.t.} \quad \sum_i c_i x_i \le B,\;\; x_i \in \{0,1\}

B is the token budget you allocate to memory, not the model's advertised window. Every memory system is solving this approximately and none of them will tell you what B is — set it yourself.

Memory systems

Third-party stores for facts, sessions and entities.

Graphiti

Bi-temporal graph where edges carry valid-from and invalid-at, so facts expire

Temporal invalidation is the clearest differentiator against native memory tools.

temporalgraph

Apache-2.0open sourceproduction viableself-host

Mem0

Extracts facts into a vector, graph and KV store scoped by user, session or agent

fact-extractionpopular

Apache-2.0open coreproduction viableboth

Zep

Async summarization and entity extraction layered over Graphiti

summarizationservice

Apache-2.0open coreproduction viableboth

Cognee

ECL pipeline building an ontology-typed graph alongside a vector index

ontologygraph

Apache-2.0open sourceearlyself-host

Hindsight

Retain, recall and reflect over memory banks with fact extraction and an MCP server

mcpreflection

MITopen sourceearlyself-host

LangMem

Extracts and updates memories into a LangGraph store

Recent activity is largely dependency bumps, with no tagged releases.

langgraphcheck-before-adopting

slowingMITopen sourceearlyself-host

Letta (ex-MemGPT)

LLM-as-OS paging between main context, recall and archival stores

Core repo took ~5 commits in three months; development moved to sibling repos. Still widely recommended in roundups.

pagingcheck-before-adopting

slowingApache-2.0open coreearlyboth

MemOS

OS-style abstraction over plaintext, activation (KV) and parametric memory

research-derived

Apache-2.0open sourceearlyself-host

Supermemory

Memory API and RAG over user context, with a router SDK and MCP

mcpapi

MIT clientopen coreearlyboth

Native platform state

Memory and state primitives shipped by model vendors.

Anthropic context editing

Server-side clearing of stale tool results and thinking blocks before token counting

nativeanthropicpruning

Commercialcommercialproduction viablesaas

Anthropic memory tool

Claude issues CRUD against a /memories directory; you own the storage backend

Stores nothing itself. Budget for path-traversal defenses, per-tenant isolation and retention policy.

nativeanthropic

Commercialcommercialproduction viablesaas

Gemini context caching

Explicit cached-content handles or implicit auto-cache on repeated prefixes

nativegoogle

Commercialcommercialproduction viablesaas

OpenAI Responses API state

Server-stored response items chained via previous_response_id, 30-day retention

nativeopenai

Commercialcommercialproduction viablesaas