MemoryAtlas

Tiered context (L0→L2) for token savings; unifies memory + resources + skills as a filesystem. Built by ByteDance subsidiary Volcengine; native hook-based integrations for OpenClaw, Hermes, Claude Code, and Codex CLI; MCP endpoint for generic MCP clients (Cursor, OpenCode, ChatGPT, etc.).

Deep diveOpenViking deep dive: what it's actually like to run agent memory as a markdown filesystemRead →
Storage
Self-hosted server by default, installed via `pip`/`uv tool install`/`pipx install openviking` or a Docker Compose image (`ghcr.io/volcengine/openviking`) — official docs (2026-07) now default to these prebuilt packages and explicitly say not to assume a Go/Rust/C++/CMake toolchain is required unless a from-source build is confirmed necessary. Requires a configured VLM (vision-language model) AND embedding model — there is no embedding-only setup. A managed alternative, OpenViking Personal (VikingDB-backed), launched in 2026 for teams that don't want to run the server themselves.
Retrieval
Hierarchical filesystem traversal over a tiered L0→L2 markdown structure combined with semantic search; retrieval trajectory is visualized for debugging.
Self-host
Self-host: moderate
License
AGPL-3.0
Pricing
Open-source core (AGPL-3.0) is free to self-host. OpenViking Personal, a hosted tier built on VikingDB, launched in 2026 with a free trial for up to 50 files; pricing beyond the trial is not publicly listed. · Free + paid
GitHub stars
28,159
Last release
2026-08-06
Last commit
2026-08-10
First catalogued
2026-06-28

Strengths

  • Up to 91% token reduction vs native agent memory (OpenClaw on LoCoMo benchmark, self-reported)
  • Deterministic, inspectable retrieval via filesystem ls/find
  • Self-evolving: automatically compresses sessions and extracts long-term memory
  • Unifies memory, resources, and skills under one filesystem paradigm
  • Genuine adoption signal beyond stars: 173+ contributors, sustained weekly commit activity since launch, and named users including ByteDance, Meituan, Renmin University of China, and CUHK-Shenzhen

Watch out

  • Relicensed from Apache-2.0 to AGPLv3 in March 2026, explicitly to require closed-source commercial derivatives to open-source — community discussion drew comparisons to the Redis-to-Valkey and Elasticsearch-to-OpenSearch forks that followed similar moves; only the Rust CLI and examples were carved back out as Apache-2.0
  • Default self-host requires both a VLM (vision-language model) and an embedding model configured — more setup surface than a vector-DB-only memory layer. A Go/Rust/C++/CMake toolchain is only needed for an optional from-source build; the documented default is pip/uv/pipx or Docker (corrected 2026-07-23 — an earlier version of this card overstated the toolchain as a default requirement)
  • Token savings vary by agent: 91% for OpenClaw, 63% for Claude Code, 34% for Hermes (all self-reported)
  • Token reduction for Hermes (34%) is the smallest of the three benchmarked agents, because Hermes already has its own memory compression — net gain is still a 2.48x accuracy improvement
  • No documented supersession mechanism for contradictory memories beyond a manual 'forget' tool — long-running deployments should plan for how stale or conflicting facts get reconciled

Best for

  • Token-lean, inspectable filesystem-based context (no vector DB required)
  • OpenClaw / Hermes / Claude Code / Codex CLI agent setups unifying memory + resources + skills
  • Teams that need observable, debuggable retrieval paths

How it integrates

Benchmark results

No sourced results yet.

Sources

Last verified 2026-08-01 · updated by refresh-framework-cards