How Merso compares

Eleven memory systems for AI agents, side by side with what Merso is being built to do. Merso is in design, so its row is the goal, marked with the stage each part ships in. Everything else works today.

Based on each project's public documentation, checked on 26 October 2026.

Where Merso is headed

Each system is a point. Across: how well it manages memory after saving. Up: how widely it captures work and reaches your tools. Merso aims for the top-right corner, where both are true.

Merso (goal): manages memory 6 of 6, reach 4 of 4Merso (goal)claude-mem: manages memory 0.5 of 6, reach 3 of 4; 5 more not documentedclaude-memMem0: manages memory 1 of 6, reach 3 of 4; 3 more not documentedMem0Cognee: manages memory 1.5 of 6, reach 2.5 of 4; 4 more not documentedCogneeGraphiti / Zep: manages memory 3 of 6, reach 2 of 4; 4 more not documentedGraphiti / ZepLetta: manages memory 1.5 of 6, reach 2 of 4; 4 more not documentedLettamemsearch: manages memory 2 of 6, reach 2.5 of 4; 4 more not documentedmemsearchSupermemory: manages memory 0.5 of 6, reach 3.5 of 4; 5 more not documentedSupermemoryBasic Memory: manages memory 0 of 6, reach 3.5 of 4; 6 more not documentedBasic MemoryPydantic AI: manages memory 1.5 of 6, reach 0.5 of 4; 3 more not documentedPydantic AIClaude Code: manages memory 1.5 of 6, reach 1 of 4; 3 more not documentedClaude Codedeveloper-os: manages memory 1.5 of 6, reach 2 of 4; 4 more not documenteddeveloper-os
Points count what each project documents: yes is 1, partly is 0.5. Features a project doesn't document count as zero, so systems with thinner documentation may sit lower than they deserve. The criteria are the ones Merso cares about, so this is a map of Merso's goals, not a general ranking. The full detail is in the tables below.

What stands out

Everyone has one piece

Graphiti handles time and duplicates but doesn't capture work from your tools. Basic Memory is readable and linked but doesn't tidy itself. claude-mem captures and serves context well but doesn't track when facts change. Merso's goal is all of it in one memory.

Two things we haven't seen anywhere

Management rules kept in a plain file that the reconciler follows. Open loops that close themselves, with the session or commit that closed them as evidence.

Where others are ahead

They all work today. claude-mem has a huge community, Mem0 and Zep publish benchmark results, and Graphiti's model of time is the one Merso borrows. Merso still has to build its goals and prove them with its own evaluation.

Capture and reach

Does memory build itself, and does it follow you into every tool you work in?

SystemCaptures work on its ownWorks across toolsOne memory for all projectsWorks from chat apps
Mersothe goal
Goal, stage 1: Stage 1hooks and adapters
Goal, stage 1: Stage 1Claude Code and Codex first; Pi, Flue and Pydantic AI in stage 2
Goal, stage 1Stage 1
Goal, stage 1: Stage 1claude.ai and ChatGPT over MCP
claude-mem
Yes: hooks
Yes: Claude Code, Codex, Cursor, OpenCode and more
Partly: current project by default, "all" setting
Partly: Claude Desktop; hosted server in beta
Mem0
Yes: plugin
Yes: API and MCP
Partly: plugin keeps memory per repository
Partly: MCP
Cognee
Yes: plugin
Partly: Claude Code and Codex
Yes: shared dataset by default
Not documented
Graphiti / Zep
Not documented: a library; no capture from coding tools documented
Partly: reference MCP server
Yes: groups
Partly: MCP
LettaClaude Subconscious
Yes: reads every session
No, per its own documentation: built for Claude Code
Yes: by default
Not documented
memsearch
Yes: after every reply
Yes: Claude Code, Codex, OpenClaw, OpenCode
Partly: files per project
Not documented
Supermemory
Yes: at session end
Yes: Claude Code, Codex, OpenCode
Partly: per repository plus personal memory
Yes: hosted MCP
Basic Memory
Partly: mostly on request
Yes: MCP
Yes
Yes: MCP
Pydantic AIHarness Memory
No, per its own documentation: the agent writes through tools
No, per its own documentation: a Pydantic AI capability
Partly: scopes set by the app
Not documented
Claude Codebuilt-in memory
Partly: preferences and corrections, not a work log
No, per its own documentation: Claude Code only
Partly: per repository; can point to one folder
No, per its own documentation: Claude Code only
developer-os
Partly: nudge after 30 tool calls
Partly: Claude Code; Codex experimental
Yes
Not documented

Managing memory

What happens after something is saved: updates, duplicates, forgetting, sources and rules. This is what Merso is built around.

SystemFacts with timeMerges duplicatesForgetsShows where a fact came fromManagement rules in a fileTracks open loops
Mersothe goal
Goal, stage 2: Stage 2valid from, valid to, retracted
Goal, stage 2: Stage 2aliases, merged by the reconciler
Goal, stage 2: Stage 2expiry and dormancy
Goal, stage 2: Stage 2session and tool behind every fact
Goal, stage 2: Stage 2constitution.md
Goal, stage 1: Stage 1shown in the briefing; close themselves in stage 2
claude-mem
Not documented
Not documented
Not documented: compresses at capture time
Partly: observations cited by ID
Not documented: settings for model and context
Not documented
Mem0
Partly: add-only; picks the current fact by date at search time
Partly: links entity names
No, per its own documentation: add-only: memories accumulate
Not documented
Not documented
Not documented
Cognee
Not documented
Partly: skips unchanged content
Partly: manual forget command
Partly: cites the memories it used
Not documented
Not documented
Graphiti / Zep
Yes: invalidates, never deletes
Yes
Not documented
Yes: every fact points to its source
Not documented
Not documented
LettaClaude Subconscious
Not documented
Partly: agent rewrites memory blocks
Partly: agent rewrites memory blocks
Not documented
Not documented
Partly: pending_items block
memsearch
Not documented
Partly: skips unchanged content
Partly: compact command
Yes: links back to the original session
Not documented
Not documented
Supermemory
Not documented
Not documented
Not documented
Not documented
Partly: settings for what to capture and recall
Not documented
Basic Memory
Not documented
Not documented
Not documented
Not documented
Not documented
Not documented
Pydantic AIHarness Memory
Not documented
Partly: a hint to the model
Partly: the model deletes files
No, per its own documentation: no source citations, per its docs
Partly: configurable guidance
Not documented
Claude Codebuilt-in memory
Not documented
Partly: Claude edits its own notes
Partly: Claude deletes wrong notes
Not documented
Partly: instructions in CLAUDE.md
Not documented
developer-os
Not documented
Partly: human review
Partly: gardening
Not documented
Partly: agent rules and human approval
Not documented

Context, search and ownership

How much goes into the context window, how things are found, whether a person can read it, and where the data lives.

SystemShort start, the rest on demandSearch by meaningKnowledge graphReadable and editableWhere the data livesLicenceMaturity
Mersothe goal
Goal, stage 1: Stage 1a briefing of about 60 lines, details over MCP
Goal, stage 1: Stage 1words, meaning and graph neighbours
Goal, stage 2: Stage 2things and relations with time, in SQLite
Goal, stage 1: Stage 1Markdown views, export to Obsidian
your Cloudflare account, or managedApache-2.0in design
claude-mem
Yes: index, then timeline, then details
Yes
Not documented
Partly: web viewer; data in a database
local, optional cloud syncApache-2.0≈103k stars
Mem0
Partly: up to 5 memories at start
Yes
Partly: entity linking
Not documented
cloud (plugin); self-hostable libraryApache-2.0≈68k stars
Cognee
Partly: recall on every prompt, up to 9,000 characters
Yes
Yes
Partly: graph visualisation
local or cloudApache-2.0≈33k stars
Graphiti / Zep
Not documented
Yes: plus graph traversal
Yes: temporal
Not documented: stored in a graph database
your graph database (Neo4j, FalkorDB)Apache-2.0≈32k stars
LettaClaude Subconscious
Partly: short hints before each prompt
Not documented
Not documented
Partly: text blocks
Letta cloud; self-hosting possibleMIT≈2.9k stars, labelled a demo
memsearch
Yes: search, expand, original
Yes: local embeddings
Not documented
Yes: Markdown
localMIT≈2.8k stars
Supermemory
Yes: profile at start, recall per prompt
Yes
Not documented
Not documented
Supermemory cloudMIT (plugin)≈2.8k stars
Basic Memory
Partly: briefing via plugin
Yes
Yes: links and relations
Yes: Markdown, Obsidian
local or cloudAGPL-3.0≈4.3k stars
Pydantic AIHarness Memory
Yes: about 2,000 tokens at start
No, per its own documentation: literal text search
Not documented
Yes: Markdown
your app (files, SQLite, Postgres)MITofficial library, 0.x
Claude Codebuilt-in memory
Yes: 200 lines at start
Not documented
Not documented
Yes: Markdown
localpart of Claude Codebuilt in
developer-os
Partly: searches before work
Not documented: text search (ripgrep)
Partly: Obsidian links
Yes: Obsidian
localMITearly

Where the ideas come from

Merso borrows openly. This is what comes from where.

IdeaFrom
Facts with a validity window; invalidate instead of deleteGraphiti
A short index at start, details on demandclaude-mem, Pydantic AI, Claude Code
A daily journal in Markdown that links back to the original sessionmemsearch
Several memory scopes; the app, not the model, decides whose memory it isPydantic AI
Fact categoriesMem0
"Remember this" with human reviewdeveloper-os
Memory treated as untrusted contentPydantic AI
Management rules as a file; open loops that close themselves with evidencenew in Merso

Help build it

Merso is open source under Apache-2.0 and designed in public. Stage 1 comes next, and these are the parts where help matters most.

  • Adapters for more tools

    Pi, Flue, Pydantic AI, Cursor, Gemini CLI. Each one is a thin layer that sends finished sessions.

  • The memory service

    Cloudflare Workers, a Durable Object with SQLite for each memory owner, and an MCP server.

  • Reconciler rules

    What to keep, merge and forget, written down in constitution.md.

  • Evaluation

    Real questions with known answers, to measure whether memory actually helps.

The repository opens with v0.1. Until then, the roadmap shows what ships in which stage.

Comparison as of 26 October 2026. These projects move fast, so some details may have changed since.