You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Working + long-term memory server over Redis: semantic search, summarization, topic/entity extraction, exposed via REST/MCP. Backed and maintained by Redis Inc.
Inclusion criteria
Agent memory is a primary concern of the project, not an afterthought
Publicly accessible (OSS, or has a public API)
Reasonable maintenance signal in the last 6 months — verify at vetting (corporate-maintained — strongest maintenance signal in the candidate set)
Permissively licensed, or open-core included via its OSS edition only
Anything else reviewers should know?
Instantly familiar to the benchmark's audience (the 'redis of X' mental model). Runs self-hosted against OSS Redis. Verify its extraction/embedding model configuration is injectable so the shared-models rule holds.
Would you be able to contribute the adapter yourself?
Maintainer-filed — up for grabs; guidance available. Adapter lands as benchmarks/<system>/ per the README's 'Adding a memory system'.
Not benchmarked. The one box this issue left open — "Reasonable maintenance signal in the last 6 months — verify at vetting" — fails on verification, and it fails specifically for the edition the inclusion criteria would admit.
It is no longer actively maintained and is not the supported production path.
The supported product is Redis Agent Memory in Redis Iris — a managed Redis Cloud service reached with an endpoint, a Store ID, and an API key. Since the relocation, commits touching V0/ have been landing-page and benchmark changes, not server work.
So this issue's premise — "Backed and maintained by Redis Inc. … corporate-maintained, strongest maintenance signal in the candidate set" — no longer holds for the OSS edition.
Why that decides it
The criteria say open-core projects are included via their OSS edition only, and that hosted SaaS without an OSS counterpart is out of scope. An OSS counterpart does exist, so this is not a clean SaaS rejection. But it means a row labeled "Redis Agent Memory" would measure an archived research artifact, while Redis ships something different and hosted under that name. That is a misleading row, and the cost of a full LongMemEval sweep is real. Not worth it for an archive.
Worth noting alongside: Redis now publishes their own LongMemEval harness in the same repository, evaluating the shipping REST API.
What did check out
Recording this so reversing the call is cheap — nothing here is the blocker:
Check
Result
License
Apache-2.0 (this issue said "MIT (verify current)" — it is Apache-2.0, Redis Inc.)
Self-hostable without vendor credentials
Yes — DISABLE_AUTH=true, published image redislabs/agent-memory-server
Extraction/embedding models injectable (the shared-models rule raised above)
Yes — GENERATION_MODEL / EMBEDDING_MODEL env vars
Compose-able
Yes — API + redis/redis-stack-server, health at /v1/health
Client SDK
agent-memory-client
One design note for whoever picks this up later: extraction to long-term memory runs in a background task-worker, so an adapter must wait for that queue to drain before searching. Ingest-then-query-immediately reads an empty index.
Reopen if
V0/ returns to active maintenance, or
Redis ships a self-hostable edition of the Iris product.
Closing as not planned; the entry stays in the README's candidates table pointing here, which is where verdicts are meant to live.
Activity
RonsenbergVI commentedon Sep 15, 2026
Vetting verdict: declined
Not benchmarked. The one box this issue left open — "Reasonable maintenance signal in the last 6 months — verify at vetting" — fails on verification, and it fails specifically for the edition the inclusion criteria would admit.
What changed since this was filed
Redis restructured the repository in July 2026. The self-hostable server now lives under
V0/and is labeled by its own README as:The supported product is Redis Agent Memory in Redis Iris — a managed Redis Cloud service reached with an endpoint, a Store ID, and an API key. Since the relocation, commits touching
V0/have been landing-page and benchmark changes, not server work.So this issue's premise — "Backed and maintained by Redis Inc. … corporate-maintained, strongest maintenance signal in the candidate set" — no longer holds for the OSS edition.
Why that decides it
The criteria say open-core projects are included via their OSS edition only, and that hosted SaaS without an OSS counterpart is out of scope. An OSS counterpart does exist, so this is not a clean SaaS rejection. But it means a row labeled "Redis Agent Memory" would measure an archived research artifact, while Redis ships something different and hosted under that name. That is a misleading row, and the cost of a full LongMemEval sweep is real. Not worth it for an archive.
Worth noting alongside: Redis now publishes their own LongMemEval harness in the same repository, evaluating the shipping REST API.
What did check out
Recording this so reversing the call is cheap — nothing here is the blocker:
DISABLE_AUTH=true, published imageredislabs/agent-memory-serverGENERATION_MODEL/EMBEDDING_MODELenv varsredis/redis-stack-server, health at/v1/healthagent-memory-clientOne design note for whoever picks this up later: extraction to long-term memory runs in a background
task-worker, so an adapter must wait for that queue to drain before searching. Ingest-then-query-immediately reads an empty index.Reopen if
V0/returns to active maintenance, orClosing as not planned; the entry stays in the README's candidates table pointing here, which is where verdicts are meant to live.