The shortlist
An assistant may not retain the project context you need between sessions. A memory layer stores what the agent learns and brings the relevant parts back later. The options range from a library you build into your own agent to plain files your coding assistant can read.
Documentation-based starting points, not hands-on benchmarks. Check each option's fit, setup and costs below. Compare them with the same worksheet.
68k stars · Costs not checked · Python
A memory engine for applications, available as a library or self-hosted server alongside a managed offering. Check model, embedding and storage configuration; self-hosted defaults can still call hosted providers.
Capability source ↗— You only need simple conversation history in a single session.
4.5k stars · Costs not checked · Python
A Markdown-based knowledge store with an MCP interface. Inspect and edit the notes, then verify your specific client's connection and whether you selected local or cloud storage.
Capability source ↗— Your organisation can't adopt AGPL-3.0 software.
7.7k stars · Costs not checked · Go
A Go-based project-memory tool with SQLite search, MCP, an HTTP API and command-line interfaces. Verify the selected client's integration, storage location and access boundaries.
Capability source ↗Costs not checked
Maintain inspectable Markdown context and source-linked claims for coding agents. Source hashes can flag changes, while model-assisted reconciliation and human review still determine corrections. Its evidence-graph experiment is not proof of universal speed or factual accuracy.
Capability source ↗— You need verified factual correctness, zero-maintenance memory or guaranteed offline generation by default.
Choose what persists: conversational facts, source-linked notes or a compiled repository wiki. Retention, retrieval ranking, source freshness and correction are different controls. An evidence graph can expose relationships and stale sources without proving the claims; compilation and update costs belong in the comparison. Keep a path to inspect, correct and delete memory, and verify where model calls send the source material. Try a correction, a deletion and a second user or project: confirm that removed material is no longer retrieved and facts do not leak between contexts. Inspect embeddings, logs and backups. Saved memory is a retrieval input, not proof of correctness. These are proposed checks, not completed vendor tests.
Separate retrieval, persistent memory and current context →
Check the linked product sources, supported languages, regional availability, permissions and current costs for your actual setup.
Google Analytics helps us understand visits. It stays off unless you allow it. Your choice is remembered for up to 180 days. Privacy policy
