Bottom line
Inference, low confidence: adopt and wrap a maintained memory layer (Mem0's open-source SDK, or Graphiti for a temporal knowledge graph) behind a thin interface of your own, rather than designing a new one. Do not rely on any published memory benchmark score: the scores are self-reported, sometimes from a managed platform or an ensemble of runs, and graded by LLM judges. Re-run LoCoMo or LongMemEval yourself with the dataset pinned by hash and the judge model pinned by name and date before you trust a number. Confidence is low because the evidence pack itself flags adoption evidence as thin.
Adopt, do not rebuild
License checked on 2026-09-23; non-permissive licenses are named in the text.
- mem0ai/mem0 Apache-2.0: memory layer with Python and TypeScript SDKs, per-agent plugins, a server, and a test file per LLM, embedding and vector-store backend. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:7
Ships Python + TS SDKs, per-agent-plugin SDKs
; ecosystem/pickup/_evidence/agent-memory.md:21tests/llms/ (17 files: test_openai, test_anthropic, test_ollama, ...)
) - getzep/graphiti Apache-2.0: temporal knowledge-graph library for agent memory with separate server and MCP packages, gated by lint, typecheck and CodeQL workflows. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:11
Temporal knowledge-graph library for agent memory (Zep's core)
; ecosystem/pickup/_evidence/agent-memory.md:11strict CI (lint, typecheck, codeql, Socket firewall, pinned action SHAs)
) - letta-ai/letta-code Apache-2.0: where Letta (the MemGPT successor) is now developed; the older letta-ai/letta repo is a landing page. Watch-level evidence only: existence, activity and CI file names; its CI, tests and code were not read. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:9
active code moved to letta-ai/letta-code
; ecosystem/pickup/_evidence/agent-memory.md:10Where Letta is actually developed
; ecosystem/pickup/_evidence/agent-memory.md:10CI includes nightly-update-smoke.yml, telegram-live-smoke.yml, per-agent-watcher workflows
) - SuperMemoryAI/supermemory MIT: a monorepo memory product with per-SDK packages. Watch-level here: existence, layout and file names only; its CI and code were not read. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:8
per-SDK Python packages (openai-sdk, pipecat, cartesia...)
)
Copy these practices
- Say which system and which run produced a score: Mem0's README states its headline numbers come from its managed platform, not the open-source SDK; Supermemory's README discloses that its ~99% headline is an ensemble and 85.86% is the single-pass figure. Starter kit: Claim discipline enforced: README claims cross-checked against proofs. [Maintainer claim] (ecosystem/pickup/_evidence/agent-memory.md:37
reflect Mem0's managed platform, which includes proprietary optimizations not available in the open-source SDK
; ecosystem/pickup/_evidence/agent-memory.md:36the single-pass 85.86% is the comparable number
) - State four fields on every benchmark number: single-pass or ensemble, judge model and date, managed or open-source, dataset hash (the companion's proposed GATE-MEM-1). Starter kit: Receipts carry provenance; generation binding. [Inference] (ecosystem/pickup/pickup-agent-memory.md:148
every published benchmark number states four fields
) - Re-run the public benchmarks, do not quote them: LoCoMo ships its dataset file and per-provider judge scripts; LongMemEval splits generation, retrieval and evaluation into separate stages. Starter kit: Incumbent oracle pinned before implementation. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:13
Ships data/locomo10.json, task_eval/
; ecosystem/pickup/_evidence/agent-memory.md:28Eval pipeline split into generation → retrieval → evaluation stages
) - Keep benchmark harnesses in the repo, with config, ingestion and evaluation split: getzep/zep keeps
benchmarks/locomo/andbenchmarks/longmemeval/plus a checkpointed, retriedzep-eval-harness/. Do not adopt Zep itself as a self-hosted memory server: its README now describes the repo as Zep Cloud examples and says the Community Edition is no longer supported, with its code moved tolegacy/(licenses.tsv read, 2026-09-23). The harness code was not read. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:27Benchmark harness in-product: config-yaml + Makefile + tests subdir per benchmark
; stack/licenses.tsv:29Zep Community Edition is no longer supported. Its code has been moved to the legacy/ folder
) - Integration tests off by default, database drivers switched off by environment variable: Graphiti's
unit_tests.yml. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:22DISABLE_NEPTUNE=1 DISABLE_NEO4J=1 DISABLE_FALKORDB=1 DISABLE_KUZU=1
) - A version bump fails CI without a changelog entry: Mem0's
ci.ymljobchangelog_check; a fresh read on 2026-09-23 of https://raw.githubusercontent.com/mem0ai/mem0/main/.github/workflows/ci.yml showed the job erroring whenpyproject.toml's version changes without an edit todocs/changelog/sdk.mdx. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:24Changelog enforcement keyed to version bump
) - Pin CI actions to SHAs and firewall dependencies: Graphiti's workflows plus
socket-firewall-connectivity.yml. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/agent-memory.md:30Supply-chain hygiene in CI: pinned action SHAs, dependency firewall
)
Build only if
- Every remembered item must carry verifiable origin and your candidate incumbent cannot record it. The evidence pack did not examine incumbents for per-item provenance, so check yours first; local-only operation alone is not a reason to build, since Mem0's harness ships a local Ollama provider config. All adopted projects here are Apache-2.0 or MIT, so license is not a reason either. The FrankenSuite substrate for provenance is unproven, and the companion's own starting point for such a build is small: one local SQLite store with a pinned embedding backend. [Verified] [Inference] (stack/licenses.tsv:52
mem0ai/mem0 Apache-2.0 permissive
; stack/licenses.tsv:28getzep/graphiti Apache-2.0 permissive
; stack/licenses.tsv:47letta-ai/letta-code Apache-2.0 permissive
; stack/licenses.tsv:88SuperMemoryAI/supermemory MIT permissive
; ecosystem/pickup/_evidence/agent-memory.md:15configs/{azure-openai,ollama,openai}.yaml
; packets/franken_agent_detection-assessment.md:190as a substrate for portable agent memory — unoccupied and unproven
; ecosystem/pickup/pickup-agent-memory.md:34A single local memory store (SQLite + a pinned embedding
)
Where FrankenSuite touches this
- None of the FrankenSuite packets implements a memory layer. Two come close. franken_agent_detection normalizes 32 coding agents' session transcripts into one schema with per-artifact origin fields and feeds the maintainer's CASS session-search tool (the workload claim is the maintainer's); TRL 6, NODUS ring Pilot. [Verified] [Maintainer claim] [Inference] (packets/franken_agent_detection-assessment.md:21
normalizes 32 agents' session transcripts into one stable schema
; packets/franken_agent_detection-assessment.md:22discovers sessions from 26 agent harnesses through this crate
; packets/franken_agent_detection-assessment.md:24(TRL 6 — see §4.9). Release artifacts exist
) - Its provenance records paths, not proof: no content hash or signature yet. It is the one FrankenSuite repo under plain MIT with no rider. [Verified] (packets/franken_agent_detection-assessment.md:15
carrying the only plain-MIT license in the FrankenSuite
; packets/franken_agent_detection-assessment.md:213What it does not record is attestation: no content hash, no producer signature
; packets/franken_agent_detection-assessment.md:236Quoted scope: none — 21-line plain MIT
) - frankengraphdb targets agent-memory and GraphRAG workloads as a graph database and names FalkorDB and Kùzu's forks, which are backends Graphiti's CI switches off in unit tests, as rivals; what ships is an embedded store with a bounded read-only GQL slice, no server and no CLI; TRL 4, NODUS ring Explore. frankensqlite's packet also names an embedded agent-memory store among its credible futures. [Verified] [Maintainer claim] [Inference] (packets/frankengraphdb-assessment.md:59
agent-memory and GraphRAG workloads needing transactional, time-travelable, branchable graph state
; packets/frankengraphdb-assessment.md:151FalkorDB (the agent-memory rival)
; ecosystem/pickup/_evidence/agent-memory.md:22DISABLE_FALKORDB=1 DISABLE_KUZU=1
; packets/frankengraphdb-assessment.md:152Kùzu (archived) and its forks
; packets/frankengraphdb-assessment.md:23an embedded Rust composition crate with a bounded GQL read slice
; packets/frankengraphdb-assessment.md:23there is no fgdbd server and no installable fgdb CLI as products
; packets/frankengraphdb-assessment.md:24TRL 4 — see §4.9). Substantive-but-unproven is the textbook Explore case
; packets/frankensqlite-assessment.md:168the project's credible futures — embedded agent-memory store
)
What we cannot say
- Who uses these libraries in production: the pack deliberately cited no adopters (ecosystem/pickup/_evidence/agent-memory.md:35
Adopter claims are thin in this pack.
). - Whether any published score holds: Supermemory's README is promotional and Mem0's headline numbers are from its managed platform (ecosystem/pickup/_evidence/agent-memory.md:36
Supermemory's README is self-promotional.
; ecosystem/pickup/_evidence/agent-memory.md:37Mem0's README is honest about the platform gap
). - Whether the benchmarks still work as oracles: LoCoMo is static, and its judge models may no longer resolve (ecosystem/pickup/_evidence/agent-memory.md:38
LoCoMo repo is static (last push 2024-08-13)
; ecosystem/pickup/pickup-agent-memory.md:319Judge-model survivability
). A widely used LongMemEval mirror is not verified hash-identical to upstream (ecosystem/pickup/pickup-agent-memory.md:323-324Community mirror fidelity: is the widely-used HF mirror
). - Letta's star count belongs to a landing page, so it says nothing about letta-code (ecosystem/pickup/_evidence/agent-memory.md:40
Star-count trap on Letta
). - Whether a memory layer beats a no-memory baseline for your agent at all; no such comparison is in the evidence.
Revisit when
- A party other than the vendor reproduces LoCoMo or LongMemEval numbers for Mem0, Graphiti/Zep or Letta with pinned judge and dataset.
- A shared memory API or conformance suite appears across two or more vendors.
- franken_agent_detection ships signed, hash-carrying provenance envelopes and a second independent consumer.
Sources
- Pickup companion: ecosystem/pickup/pickup-agent-memory.md, the charter seed, oracle candidates, and first claim registry for this area.
- Evidence pack: ecosystem/pickup/_evidence/agent-memory.md, the repositories checked against the GitHub API, with dates.
- Verdict source: stack/agent-memory.md, written under stack/METHOD.md.
Starting a build here?
Begin from the pickup companion for Agent memory systems: it carries the charter seed, the oracle candidates and how to verify them, and the first claims to register, so a new project starts from evidence instead of from scratch.