Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Pick a memory server for your coding agent

Written November 7, 2026

Your agent forgets everything between sessions. Mem0, Zep, Supermemory, Letta, Cognee — what each stores, what it costs, and whether you can read it yourself.

The problem

A coding agent starts every session knowing nothing about your project beyond what is in the prompt. The facts it learned last week — which module is fragile, what the team decided, why a workaround exists — are gone unless something stored them. A memory server is that something: a store the agent writes to and reads back through, usually over MCP. The field is crowded, every vendor claims the top benchmark score, and the two comparisons below were both written by sellers in the field. Read them for the checkable facts, not for the ranking.

Two comparisons, and who wrote them

The main comparison is InfraNodus's Agent Memory Tools Compared, dated October 24, 2026, covering ten tools. InfraNodus sells a memory product of its own, and the comparison ends by recommending it for the use case it is built for — a readable graph you can audit. Its prices, licenses, and limits are stated as taken from each vendor's own pricing page, documentation, or GitHub repository, and the prices in the table below are from the vendor pages linked in Sources.

The second is the total-agent-memory vs-competitors matrix by ausduns, dated April 2026. It promotes its own local-first server for coding agents, and its feature table marks its own row as the best on nearly every line. Treat its verdicts the same way: useful for the shape of the field, not a neutral ranking.

What each one stores, what it costs, whether you can read it

Prices are monthly list prices from the vendor's own pricing page. Where a vendor does not state something, the table says so.

ToolWhat it storesLicenceHostingStarting priceCan you read it yourself
Mem0Vector store + BM25 keyword search, with an entity graph ("Graph Memory") on the platformApache-2.0 for the library; the managed platform adds proprietary optimizationsSelf-hosted library or Cloud PlatformFree Hobby tier (10,000 adds, 1,000 retrievals a month); paid plans from $19 a month, Pro $249Partly: the platform dashboard, or the vector store you run when self-hosting
Zep + GraphitiTemporal knowledge graph: facts with validity windows, old facts marked invalid rather than deletedGraphiti Apache-2.0; Zep's license is not publishedGraphiti self-hosted; Zep Cloud or in your own cloudFree (10,000 credits a month); Flex $125 a month; Flex Plus $375Yes for Graphiti, if you query the graph database (Neo4j, FalkorDB, Neptune)
LettaMemory blocks pinned in the context window, a vector archive, and full message historyApache-2.0Letta Cloud or your own letta serverFree $0 (3 agents, 5 MiB per memory repo); Pro $20 a month (20 agents, 50 MiB)Yes: the memory blocks are plain text you and the agent edit
CogneeKnowledge graph + vector index + relational store (SQLite, LanceDB, Ladybug by default)Apache-2.0Self-hosted, or Cognee CloudSelf-hosting is free; the pricing page lists Enterprise BYOC engagements and no Cloud list pricesThrough the databases you run
SupermemoryGraph of facts (Updates, Extends, Derives relations) + embeddingsMITHosted API, or a local Docker buildFree $0 ($5 of credits a month); Pro $19 a month ($20 of credits, 3 seats); Max $100; Scale $399Through the API and SDKs; the local Docker build exposes the store
Reference MCP memory serverSmall knowledge graph of entities, observations, and relations in a local memory.jsonl fileMITLocal only, one process per clientFreeYes: open memory.jsonl in any editor

Start the reference server with npx -y @modelcontextprotocol/server-memory. It is the baseline everything else is measured against: a file the model writes and reads back, with no structure beyond line order and no way to ask what changed or what is missing.

Where the two comparisons disagree

Do not rank by benchmark score

Every vendor leads with a benchmark number, and the numbers do not travel between setups. Mem0 reports 94.4 on LongMemEval and 92.5 on LoCoMo; Zep reports 90.2 and 94.7. An independent LongMemEval run in 2025 scored Mem0 at 49.0 and Zep at 63.8. The same benchmark, different setups, different numbers — which is why no benchmark score is a column in the table above. The ausduns matrix adds its own numbers, including a 96.2% recall figure for its own server on LongMemEval, measured by its own authors. None of these settles which store fits your agent.

The decision rule

Pick by storage model, hosting, and whether you can read your own data — not by benchmark rank:

After a month, the test is whether the agent's answers got better — and only a readable store lets you trace which learnings did that. A vector store returns the nearest entries; a graph returns the subgraph around the query; a flat file is reread whole, so its token cost grows with its length. That is the cost no monthly price shows.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Codex memory v2: import memories from other agents · Run claude-mem on your own plan, redaction on · Agent Skills has a neutral home — and still no security story