Field noteNov 16, 20265 min read1 source

Maximem releases Synap Core, an open-source self-hosted agent memory server

In one minute

Maximem has released Synap Core, an open-source, self-hosted memory server for AI agents. It turns conversations into structured memories and retrieves context across sessions; for builders, the release makes persistent memory available as a deployable component rather than only a managed service.

What changed: Maximem has released Synap Core, an open-source, self-hosted server for agent memory.

What happened

Maximem describes Synap Core as the self-hosted edition of its Synap memory engine. The GitHub repository identifies the project under the Apache 2.0 license and presents it as a server for memory across conversations, sessions and devices. A Docker Compose stack is the stated deployment path: the README says one docker compose up starts the service, with the data staying on the user's disk.

The system's stated approach goes beyond storing and searching text chunks. Synap extracts facts, preferences, events and episodes from conversations; resolves the entities they concern; tracks temporal information; and retrieves context when an agent asks for it. The repository describes vector and graph retrieval as part of the engine. Maximem also says the service replaces information that is no longer true, positioning memory as an evolving representation rather than an append-only transcript store.

The README reports a LongMemEval score of 83.6% for Synap Core and 95.6% for Synap Cloud, the managed service using the same architecture. Maximem says both figures use the same answer model and judge, and links to results and an evaluation harness. The repository presents these as its own reported measurements. They are useful for understanding the company's comparison, but do not by themselves establish how the systems perform under other evaluation choices or production workloads.

Core speaks the same API as Synap Cloud, according to Maximem, so SDKs and framework integrations work with both. The company says migration to Cloud requires changing the base URL and adding an API key, without changing application code. The page lists integrations for frameworks including LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, Google ADK, OpenAI Agents SDK, Semantic Kernel, Pydantic AI and others. It says Core calls out to the team's model provider and downloads its models on first start; calls to Synap Cloud and Jev are optional. The README links to a list of outbound calls.

What the archive adds

New Runtime's earlier coverage described Tencent's Hy-Memory release on 2026-10-21 as an inspectable software stack for persistent agent memory. Synap Core is a new event in the same emerging category: a separate project offering a self-hosted memory engine, with its own API and retrieval design.

Trend signal

This supports the direction captured by Company memory needs write and correction loops: Synap's stated extraction, temporal tracking and replacement of stale information treat memory as something that must be updated, not merely searched. The repository describes a product design, not independent evidence that those mechanisms reliably prevent incorrect or poisoned memories in production.

Why now

The release arrives as agent-memory projects are increasingly being presented as complete software stacks rather than isolated retrieval techniques. Synap Core's shared API with Synap Cloud also makes local evaluation and a later move to the hosted service part of the product's stated deployment story.

Why it matters

Persistent memory changes an agent's operating boundary: it can carry user facts and past events across sessions, but those memories can also become stale or wrong. A self-hosted engine gives a team more direct control over where that information resides and a way to test memory behavior without beginning with a managed service. The common API may reduce migration work if a team later chooses cloud hosting, though that is Maximem's compatibility claim and should be verified against the team's actual integration.

The useful engineering question is not simply whether a memory benchmark score is high. Teams need to inspect what the system records, how it resolves conflicting or time-sensitive facts, what leaves the machine, and how applications can correct or delete stored information. Synap Core makes these implementation choices more tangible: builders can run a memory service themselves and assess its operational and governance fit alongside retrieval quality.

Limits and uncertainty

The LongMemEval figures are reported by Maximem; the available figures do not establish independent reproduction or performance on a team's own tasks. The README's comparison uses the same answer model and judge, but benchmark scores can still depend on dataset handling, retrieval settings and evaluation methodology. The repository describes extraction, entity resolution and temporal reasoning, but these capabilities do not guarantee accurate updates in every conversation. Teams should also test the operational implications of model-provider calls, first-start model downloads and any optional external services before treating the deployment as fully isolated.

Skill to build now

Run Synap Core in a disposable local environment before connecting it to a customer-facing agent. Feed it a small set of conversations containing a changed preference, a time-bound fact and two similarly named entities. Then query the agent across a new session and inspect whether the retrieved context reflects the correction rather than both versions. Use this test when considering any persistent-memory layer: it turns claims about extraction, temporal reasoning and entity resolution into behavior the team can review. Record the outbound network calls and deletion path as part of the same trial.

Reader takeaway

Synap Core makes agent memory available as a self-hosted, Apache-licensed server with a cloud-compatible API. Its benchmark numbers are a starting point, not a deployment decision: test memory updates, retrieval and data flows on your own workload before trusting it with persistent user context.

Evidence / sources

  1. 1github.com/maximem-ai/maximem_synap_core
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicAgent MemoryExplore the agent-memory topic hub.
  2. 02archiveField NotesOpen the latest editorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract