LLM Wiki — Synaptic Lattice

RightMemory

Properties1
tagsentity, repository, tool, llm-wiki, agentic, knowledge-graph

Leitura de 4 min

Overview

RightMemory (RightL) is a tree + graph memory substrate for coding agents. It is documented in the comprehensive report as a Tier 1 core implementation with typed edges and multi-agent collaboration over a vendor-neutral CLI. It is one of the strongest projects for structured relationships (typed edges) among those documented in the vault.

Key Facts

  • Architecture: tree + graph memory for coding agents
  • Typed edges: explicit relationship types between memory nodes (dep:, cfg:, ver:, doc:, todo:)
  • Multi-agent: designed for collaboration across agents sharing one memory
  • Interface: vendor-neutral CLI (not locked to one IDE/agent)
  • Modes: standalone (Pydantic AI agent loop) or cli-agent (delegates to Codex/Claude CLI)
  • Part of the May 2026 second wave of LLM Wiki infrastructure (per the comprehensive report timeline)

Features

Core Architecture

  • Dual document trees — Memory (durable context) + Pursuit (live intent) in one graph with one id namespace
  • Standalone or CLI-agent modes — same command surface, different execution backends
  • Global id namespace — all headings and nodes share one addressable namespace across both trees
  • CLI runtime — rightmemory command with daemon mode (JSON-over-stdio) for any agent

Ingest & Processing

  • Transcript review — automatic candidate extraction from idle agent sessions
  • Unified update queue — candidates batched by work description, not session
  • One-hour quiet period — per-session debounce before batch consolidation

Search & Retrieval

  • Structured retrieval — graph-aware, id-based, not fuzzy embedding search
  • Model-selected output — retrieval returns deterministic Markdown, no model summaries
  • Session-persistent — exact message history persisted under .runtime/sessions/
  • Correction awareness — loads Agent Correction Memory after initial draft, not before

Knowledge Graph

  • Typed edges — 15 edge types: dep:, emb:, bak:, agg:, ver:, ext:, up:, loc:, run:, cfg:, out:, in:, doc:, todo:, rel:
  • F# detail files — heading-backed files for dense subtrees, keeping main files lean
  • Shared views — MF# (file packages) and MQ# (question endpoints) for multi-agent collaboration
  • Dreamer consolidation — automated pruning of stale, duplicated, overgrown memory

MCP / Tools

  • CLI tools — retrieve, update submit/pull/undo, dreamer, insight, prune, history, status, validate
  • Daemon mode — JSON-over-stdio for programmatic integration
  • Shared-view commands — create, revise, approve, publish, join, ask, note
  • Install profiles — named profiles for project-specific memory roots

Quality & Governance

  • Schema validation — rightmemory validate checks complete Memory/Pursuit graph + corrections.md
  • Edit-correction feedback — separate correction surface for RightMemory edits vs general agent work
  • Agent Correction Memory — curated modules (writing + design) with max 15 entries each
  • Git-native provenance — every update recorded as immutable JSON + Git diff

Integration

  • Codex CLI + Claude Code CLI — built-in delegated execution
  • 3 independent skills — memory-retriever (read-only), rightmemory-orchestrator (full), maintain-rightmemory (explicit-only)
  • Shared views — MF# file packages + MQ# question endpoints for cross-agent collaboration
  • Multi-device continuity — Git-syncable markdown across laptops, desktops, clients

Data & Storage

  • Markdown document trees — Memory.md + Pursuit.md + detail files
  • Git-backed — full version history, branching, sync
  • Update records — immutable JSON provenance per candidate batch
  • Insight logs — timestamped reflections under insight_logs/
  • No embeddings — structure-first, not vector-based

Relationships

  • “uses” knowledge-graph — typed-edge graph memory is its core substrate
  • “belongs to” agent-memory-systems — memory-first family, agent-oriented
  • “differs from” cognee — sibling graph-backed memory engine (Cognee is broader/multi-source; RightMemory is coding-agent focused)
  • “extends” llm-wiki — provides the memory layer coding agents use while building a wiki
  • “compatible with” model-context-protocol — agent-tool access pattern it can plug into

Implications

RightMemory leads on typed relationships for coding agents specifically. Where the LLM Wiki pattern uses untyped [[wikilinks]], RightMemory shows the typed-edge alternative that rohitg00’s LLM Wiki v2 also advocates. For this vault, it is a reference for what “typed edges” look like when memory is shared across multiple agents.

Sources

Visão de gráfico

Backlinks