Skip to content
GRPNR.

R7 — OSS foundation candidates for “agent question loop + knowledge write-back”

Research date: 2026-07-18. All GitHub stats pulled live via api.github.com (Bash/curl) on 2026-07-18 unless noted. Contributor counts are the Link: rel=last page count from /contributors?per_page=1 (approximate, undercounts anonymous/squashed commits but fine for ranking).

Raw GitHub API data (2026-07-18)

repo stars forks open issues license (SPDX) created last push contributors (approx) latest release
backstage/backstage 33,894 7,492 429 Apache-2.0 2020-01-24 2026-07-17 367 v1.53.0 (2026-07-14)
outline/outline 39,752 3,441 56 NOASSERTION (BSL-1.1, custom text) 2016-05-22 2026-07-18 240 v1.9.1 (2026-07-13)
requarks/wiki (Wiki.js) 28,625 3,268 190 AGPL-3.0 2016-08-16 2026-06-14 218 v2.5.314 (2026-05-01)
facebook/docusaurus 65,670 9,974 385 MIT 2017-06-20 2026-07-17 457 v3.10.2 (2026-07-10)
thomvaill/log4brains 1,506 112 57 Apache-2.0 2020-09-15 2024-12-17 10 v1.1.0 (~2 yrs old per npm)
HKUDS/LightRAG 37,798 5,320 215 MIT 2024-10-02 2026-07-18 267 v1.5.5rc1 (2026-07-13)
mem0ai/mem0 61,126 7,109 644 Apache-2.0 2023-06-20 2026-07-18 376 (cli sub-release) 2026-07-13
letta-ai/letta 23,853 2,533 48 Apache-2.0 2023-10-11 2026-07-03 139 0.16.8 (2026-05-14)
basicmachines-co/basic-memory 3,457 231 73 AGPL-3.0 2024-12-02 2026-07-18 ~30 v0.22.1 (2026-06-13)
getzep/graphiti 28,878 2,912 443 Apache-2.0 / MIT (docs say MIT) 2024-08-08 2026-07-17 44 v0.29.2 (2026-06-08)
langchain-ai/agent-inbox 1,032 142 21 MIT 2024-11-04 2026-07-15 6 none tagged (ever)
redis/agent-memory-server 295 59 38 NOASSERTION (verify terms) 2025-03-14 2026-07-16 21 server/v0.15.2 (2026-04-10)
topoteretes/cognee (bonus) 28,171 — — Apache-2.0 2023-08-16 2026-07-18 — —
supermemoryai/supermemory (bonus) 28,463 — — MIT 2024-02-27 2026-07-18 — —

Per-candidate notes

Backstage (incl. TechDocs)

  • Apache-2.0, huge project, daily commits, weekly-ish releases (v1.53.0 2026-07-14). Backstage.io, github.com/backstage/backstage.
  • Negative/TCO evidence: dev.to “Backstage Is Not Free: The Real TCO of Building vs Buying an Internal Developer Platform” (2026) — claims 2-3 senior platform engineers full-time ($300-450k/yr loaded), 3-yr TCO ~$1.52M; each major release costs 2-5 eng-days per custom plugin to re-validate. Port.io published “Backstage is dead” / “The Platform Engineering World Moved On” (2026) — vendor-authored but reflects real complaints about plugin fragmentation (each plugin has its own DB/context, no shared schema across catalog/incident/ownership data).
  • TechDocs specifically: GitHub issue #32815 “TechDocs Evolution: Addressing the Sustainability of the MkDocs Ecosystem” and #33990 (RFC to move to “Zensical”) confirm TechDocs is built on MkDocs (unmaintained since v1.6.1, Aug 2024) + Material for MkDocs (entered maintenance-only mode Nov 2025, committed only through Nov 2026). TechDocs also needs external cloud storage (S3/GCS) + a dedicated CI pipeline — not plug-and-play (roadie.io “Self-Hosting Backstage: The Real To-Do List”).
  • Verdict: this is a developer-portal/software-catalog product, not a Q&A/memory loop. High deploy complexity, real TCO, and the one plugin nearest our topic (TechDocs) sits on a foundation the Backstage community itself is worried about.

Outline

  • License is Business Source License 1.1 (GitHub shows NOASSERTION because BSL isn’t a recognized OSS SPDX id). Restriction is narrow: can’t resell it as a hosted multi-tenant “Document Service” to third parties; internal self-hosted use for one’s own org is unrestricted. Converts to Apache-2.0 automatically in Jan 2030. (github.com/outline/outline/blob/main/LICENSE; docs.getoutline.com license-restrictions page; HN thread #39012054.)
  • Very active (daily pushes, weekly releases), 240 contributors, real-time collab, has an API suitable for programmatic doc writes.
  • No agent/LLM question-loop features; it’s a human wiki. Would only ever be a destination store for the “write-back” half.

Wiki.js

  • AGPL-3.0. v2 branch still gets patch releases (v2.5.314, 2026-05-01) but the project has been trying to ship a v3 rewrite since a “Developer Preview” in Oct 2022; as of mid-2026 the official release notes still list v3 beta as “No ETA” / “under development” — a 3.5+ year stall on the project’s own stated roadmap (js.wiki release notes, Wikipedia “Wiki.js” article cross-check). This is a real stagnation signal even though v2 patches continue.

Docusaurus

  • MIT, Meta-backed, by far the largest contributor base here (457) and highest star count (65.6k), releases essentially weekly, pushed same-day as this research.
  • Purely a static-site generator, docs-as-code (git + build step). No runtime database or write API — an agent “writing back” means committing markdown + rebuilding, which is the wrong shape for a live, per-question write-back loop. Best used only as the eventual human-facing publish target, not the loop itself.

log4brains

  • github.com/thomvaill/log4brains. Apache-2.0, only 1,506 stars, 10 contributors, and its last commit/release is Dec 2024 (npm: “last published 2 years ago” as of this search) — 18+ months stale as of 2026-07-18. Niche ADR-only tool. Community commentary already flags “long pause” pattern. Effectively a one-person side project.

LightRAG (operator already runs this locally via skills)

  • HKUDS/LightRAG, MIT, EMNLP 2025 paper, 37.8k stars, 267 contributors, releases essentially weekly (rc July 13 2026), actively adding features in 2026 (OpenSearch unified storage backend, setup wizard, local embedding/reranking/storage in Docker per March 2026 changelog).
  • Already the operator’s in-place knowledge-graph RAG engine (per lightrag-* skills). Solves ingestion/retrieval + graph write-back well; does NOT solve “ask a human a question and route the answer back” — no HITL/escalation primitive.

mem0

  • mem0ai/mem0, Apache-2.0, 61.1k stars (highest of the memory-specific projects), 376 contributors, VC-backed (Mem0 Inc., USA), extremely active (daily pushes).
  • Real credibility risk: LOCOMO benchmark dispute is public and ongoing — Zep originally claimed 84% vs Mem0, Mem0 disputed the figure down to 58.44%, Zep counter-claimed 75.14%; an independent test found Mem0 accuracy degrading sharply as evidence items increase (61%→38%→25% as items go 1→3→6). Treat mem0’s own performance marketing skeptically (atlan.com “Zep vs Mem0”, multiple 2026 comparison posts).
  • API shape (add/search memories) is personalization/user-preference memory, not a question-router/HITL loop.

Letta (MemGPT)

  • letta-ai/letta, Apache-2.0, 23.9k stars, ~139 contributors, still gets commits (last push 2026-07-03) but releases have slowed (0.16.8, 2026-05-14).
  • Major maintenance-risk finding: Letta’s own 2026 blog (“Letta’s Next Phase”) states they are sunsetting server-side memory features (e.g. core_memory_replace) in favor of client-side/filesystem-native “MemFS” and consolidating the whole company around Letta Code, a coding-agent product — standalone LettaBot was archived in May 2026 and folded into Letta Code as “Channels.” The tiered-memory server architecture that would make Letta a “foundation” for this concept is being actively dismantled by its own maintainers in favor of a different product line.

basic-memory

  • basicmachines-co/basic-memory, AGPL-3.0, 3,457 stars, ~30 contributors (small team/single company “Basic Machines”), active (release 2026-06-13, push 2026-07-18).
  • Conceptually close (LLM reads/writes a local markdown knowledge graph via MCP) but positioned for single-user/Obsidian-integrated personal use, not evidenced at any multi-agent/production scale. No question-routing or HITL feature.

Zep / Graphiti

  • Zep discontinued and stopped updating its self-hosted open-source “Community Edition” in April 2025 (blog.getzep.com “Announcing a New Direction for Zep’s Open Source Strategy”) — teams that want to self-host now must run raw Graphiti + a graph DB (Neo4j/FalkorDB/Kuzu) themselves, materially more infra work than before. This is a precedent of the vendor pulling the easy self-host path out from under adopters.
  • Graphiti itself (getzep/graphiti): 28.9k stars but only 44 contributors (concentrated, single-company-driven despite popularity), MIT-licensed per Zep’s own site (GitHub shows Apache-2.0 file — verify before adopting), active (push 2026-07-17, release 2026-06-08). Same problem space as LightRAG (temporal knowledge-graph memory) — largely redundant with what’s already deployed.

langchain-ai/agent-inbox

  • github.com/langchain-ai/agent-inbox, MIT, only 1,032 stars, 6 contributors, no releases have ever been tagged, but still receives commits (2026-07-15).
  • Conceptually the closest single match to “agent asks a question, human answers via an inbox” — but it’s a thin Gmail-style UI shell tightly coupled to a LangGraph deployment + LangSmith API key (i.e., adopting it means adopting LangChain’s whole agent runtime), with no knowledge write-back of its own. Small team, zero release discipline = real bus-factor/versioning risk for a production dependency.

Redis agent-memory-server

  • redis/agent-memory-server, created 2025-03-14 (~16 months old), only 295 stars / 21 contributors, license file shows NOASSERTION (verify commercial terms explicitly — Redis Inc. relicensed core Redis away from OSS to RSALv2/SSPLv1 in 2024, so don’t assume standard OSS terms transfer to satellite repos).
  • Redis’s own docs describe it as “currently available in preview, with features and behavior subject to change.”
  • Architecturally the best conceptual match of anything found: two-tier session (short-term, TTL’d) + long-term memory with automatic background extraction/promotion from session → long-term store, exposed via REST, MCP, and Python client, pluggable backends (Redis, Pinecone, Chroma, Postgres). This is essentially the shape of a “question loop + knowledge write-back,” just not aimed at human-escalation specifically and still labeled beta by its own vendor.
  • Cognee (topoteretes/cognee): Apache-2.0, 28.2k stars, active daily; markets itself explicitly around a “remember / recall / improve / forget” memory-native API — closest philosophical match to “knowledge write-back,” worth a bake-off vs LightRAG before ruling out.
  • Supermemory: MIT, 28.5k stars, active daily; positions specifically for coding-agent memory (Claude Code / OpenCode plugins) — relevant given the operator’s Claude Code-based agent fleet, but unverified at 64-agent multi-tenant scale.
  • Memary: described in secondary sources as “lightweight… for experimentation and prototyping rather than production-scale deployments” — explicitly prototype-only per its own positioning (tryxlr8.ai / cognee.ai comparison posts, 2026).

Sources consulted (non-exhaustive, all live-checked 2026-07-18)

Direct answer

Nothing found is a drop-in “agent question loop + knowledge write-back” product. LightRAG (already running) covers the write-back/retrieval half well. The pieces that resemble the “loop” half — agent-inbox (thin, single-vendor, zero releases, 6 contributors) and Redis agent-memory-server (vendor-labeled preview, 16 months old) — are both immature and narrow. mem0/Letta/Zep-Graphiti are broader “memory layer” products carrying real strategic-pivot or benchmark-credibility risk rather than being safe foundations. Given LightRAG already solves storage/retrieval and the missing piece (route an agent’s stuck-question to a human, capture the answer, write it back) is genuinely small glue code, building that thin loop from scratch — reusing LightRAG for storage and borrowing UX patterns from agent-inbox / Redis agent-memory-server’s session→long-term promotion shape — is NOT strategically irrational. It is the cheaper, lower-risk path versus importing Backstage’s TCO, Outline/Wiki.js’s unrelated wiki surface, or a memory vendor whose roadmap or benchmark story is currently in flux.