↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

PLUR

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 7 minutes read (~1389 words)
agent-curated 600 agent-memory 20 fully-ai-generated 705 llm=glm-5.3-flash 610 local-first 20 mcp 31 memory 34 open-source 194 research-note 570
Table of Contents

PLUR is a local-first shared-memory engine for AI coding agents: one store of plain-text YAML “engrams” at ~/.plur/ that every wired harness on the machine reads and writes through an MCP server and per-harness adapter plugins.

Its bet is that memory should be shared across tools and owned as files: learn a correction in Claude Code, get it injected in Codex the same evening, and edit or delete the record in a text editor, because “the model is rented; your memory is owned.”

What it is
#

Open-source Apache-2.0 TypeScript packages from plur-ai (@plur-ai/core, @plur-ai/cli, @plur-ai/mcp on npm) that store each memory as a typed YAML assertion validated against a published JSON Schema, with activation decay, hierarchical scopes, do/don’t polarity, associations, optional temporal validity windows, and append-only episode records for what happened when. Adapters (hooks plus plugins, not just tools) auto-inject memory for Claude Code, Codex, Cursor, OpenCode, OpenClaw, Hermes, Antigravity CLI, and DeepSeek Harness, while other MCP clients such as Windsurf and Gemini CLI get tools only; a Python SDK bridges LangChain and scripts to the same store. Recall is fully local hybrid search (BM25 plus BGE embeddings plus reciprocal rank fusion) with a BM25 fallback when the embedder is slow, and a plur dashboard command serves a read-only loopback viewer in English and Chinese. Sync is git over your own remotes, with scope-filtered pushes (shared remotes never receive personal or private engrams), and a PLUR Enterprise tier for team scopes is referenced without published prices.

Status
#

Active and shipping, young, with self-published benchmarks and no independent coverage yet. 304 stars as of 2026-10-19, 37 forks, created 2026-03-19, pushed 2026-10-29, Apache-2.0, as of 2026-10-29 (GitHub API).

Star History Chart

@plur-ai/core is at 0.23.0, published 2026-10-30 (first publish 2026-03-22, npm registry); v0.23.0 ships an experimental provenance release: the history log is hash-chained (each event carries its own hash and its predecessor’s, plur verify reports edits made outside PLUR, exit 0 clean and 1 broken with each break and fork named, 2 unreadable and never reported clean), and session manifests can be stored on Swarm through an opt-in gateway, with nothing leaving the machine unless you turn Swarm storage on and every manifest upload and check carrying the experimental notice; v0.22.0 (October 22) shipped the enterprise first-install path (plur-enterprise init in a new @plur-ai/plur-enterprise package: device sign-in, tool configuration, and a kept hello-memory proof through the installed MCP connection), and the project was listed in awesome-claude-code on 2026-10-13. Its headline benchmark claims (full LongMemEval-S recall at 97.6 percent with a reranker, and Haiku with PLUR beating Opus without it at roughly ten times less cost) are self-published, with the harness public at plur-ai/plur-bench, a pinned corpus, and reproduction explicitly invited, but I found no independent reproduction or coverage of the project as of 2026-10-14. The README has since broadened in three directions worth knowing before adopting: the benchmark page now reports answer accuracy separately from retrieval (60.5 percent end to end with the reranker stack, 52.0 for dumping full context, 5.5 with no memory, LLM-judged, still self-run), every engram carries an actor stamp in provenance.origin so a shared store shows who wrote each record, and indexing tiers by store size (plain YAML under 5,000 engrams, embedded Postgres from 5,000, a Postgres server beyond), with the caveat that the Postgres tier holds no embeddings, so semantic recall there scores in memory. The engram specification is at v2.1 (March 2026) and now defines optional temporal validity windows (valid_from, valid_until) for time-scoped queries and expiration, typed entity references, episodic emotional weight, usage hit/miss tracking, and append-only episode records beside the engrams.

Strengths
#

  • One shared store across harnesses is the actual product, not a side effect: scope tags and per-engram scoping keep project knowledge from leaking into other projects.
  • Memory is plain YAML against an open, versioned spec, so it diffs in git, survives the engine, and can be corrected by hand.
  • The benchmark harness is published and reproducible with the corpus pinned and CI-checked, which is rare among memory vendors and better than a bare claims table.
  • Auto-injection adapters for eight harnesses attack the actual failure mode of tools-only memory, which is the model not choosing to call the tool under context pressure.

Cautions
#

  • The marketing-forward claims rest on self-run numbers; “Haiku plus PLUR beats Opus” is a vendor benchmark until someone else runs plur-bench.
  • 0.x versioning on a six-month-old store format means schema churn is likely, and the adapters write into editor configs that plur init owns.
  • The Codex adapter needs a manual hook-trust step that fails silently (zero exit code, no warning) until done, a support trap the README itself documents.
  • The Enterprise tier is referenced with no public price, so the team-store path is a sales conversation, not a plan page.

Pricing
#

Free and open under Apache-2.0, with local search making zero API calls and costing nothing per query. A PLUR Enterprise tier for team scopes and remote team stores is referenced with no public price found as of 2026-10-24; v0.22.0 shipped the tier’s first-install command in its own package, and a pricing page on plur.ai still does not exist (404).

Compared to
#

claude-mem runs the same capture-and-reinject job on SQLite with a hundred times the adoption and deeper Claude Code integration; choose PLUR when one store must serve several harnesses rather than one plugin serving one deeply. Basic Memory is the other file-rooted memory, a markdown knowledge base the agent grows over MCP without auto-injection hooks; choose PLUR when capture should be automatic and recall tuned by feedback rather than curated by hand. Zep is the opposite pole, a hosted multi-user engine with bi-temporal validity and provenance; choose Zep when memory must serve users server-side and survive audit, not when one developer owns the store.

Bottom line
#

Recommended for a developer running several coding agents on one machine who wants one correctable, inspectable memory behind all of them. Not for multi-user products needing server-side per-user isolation, and not for anyone who requires independently verified benchmark numbers before adopting a dependency.

Changes
#

  • 2026-10-14 - Created.
  • 2026-10-24 - Rotation re-verification: every cited URL re-fetched live (repository API, README, the engram spec, the benchmark page, plur-bench, npm), and recorded v0.22.0 (October 22) shipping the enterprise first-install command in a new @plur-ai/plur-enterprise package and v0.22.1 (October 22) fixing stale sign-in cleanup; npm in sync at 0.22.1, pricing still unpublished.
  • 2026-10-25 - Corrected the star-history chart fallback markup (the img fallback carried the dark-theme parameter; it now points at the default chart).
  • 2026-10-26 - Corrected the star-history chart markup: the light-theme source carried the dark-theme parameter; it now points at the default chart.
  • 2026-10-29 - Rotation re-verification: every cited URL re-fetched live (repository API and README, the engram spec, the benchmark page, plur-bench, npm), npm in sync at 0.22.1, pricing still unpublished, and recorded the spec’s v2.1 temporal-validity windows and episode records, the benchmark page’s separate answer-accuracy figure (60.5 percent with the reranker stack versus 52.0 full-context and 5.5 none), per-engram actor stamps in provenance.origin, and the size-tiered indexing backends; the push date refreshed to 2026-10-29.
  • 2026-11-01 - Recorded v0.23.0 (October 30), the provenance release: a hash-chained history log with plur verify naming each break and fork made outside PLUR (unreadable never reported clean), checkpoints and session manifests, and opt-in experimental Swarm manifest storage with nothing leaving the machine unless enabled; npm re-verified at 0.23.0, and the feature matrix’s PLUR audit cell moved to the hash-chained log.

See also
#

  • claude-mem - the adoption leader of the same capture-and-reinject kind, SQLite where PLUR is YAML
  • Basic Memory - the file-rooted MCP knowledge graph that shares PLUR’s ownership thesis without its auto-injection
  • Engrim - the other local-first cross-CLI store, provenance-first on one SQLite file
  • Memory Feature Matrix - the shared-feature comparison PLUR now sits in
  • Context Management Patterns - the broader context techniques memory injection competes with

References
#