Lians-ai/Lians

Lians is a substantial memory platform that goes well beyond storing text and running vector search. Its strongest feature is the way it treats memory as changing evidence. Facts can be corrected, superseded, recalled at a past point in time, traced through lineage, or erased through governed controls. Bounded recall helps prevent an agent from flooding its context with old information. The repository supports a practical local path through SQLite, semantic embeddings, MCP, and a desktop installer, while also providing a full FastAPI service with PostgreSQL, Redis, encryption, audit chains, information barriers, and tenant isolation. Python, TypeScript, Go, Java, and C SDKs broaden its usefulness, and integrations cover several agent frameworks and coding tools. Engineering practices are strong. CI tests Python against PostgreSQL and Redis, exercises every SDK, verifies migrations and API contracts, audits dependencies, scans with CodeQL and Gitleaks, builds non-root containers, and runs reproducible memory evidence gates. The benchmark documentation also distinguishes local functional checks from production and competitive results instead of overstating what has been proven.
The main weakness is repository complexity. Lians contains the core service, multiple SDK generations, desktop packaging, integrations, plugins, deployment files, commercial documentation, benchmarks, and compliance material in one large tree. This makes it difficult to identify the current supported path. For example, agentmem/sdk/python contains the current lians-sdk package at version 0.5.0, while a separate sdk/python directory still describes an older lians package at version 0.2.0. The changelog also has two Unreleased sections and explains that 0.5.0 source manifests do not yet represent a fully published release. That honesty is good, but stale package trees should be removed, archived, or labeled clearly. A smaller architecture overview showing which components are current, legacy, local-only, hosted-only, or experimental would reduce confusion. Broader end-to-end tests for the packaged desktop installers and signed or attested release artifacts would strengthen distribution trust as well. Overall, Lians is a technically serious project with careful governance and unusually strong evidence practices. Simplifying its structure and release story would make its strongest capabilities easier to adopt and evaluate.