mirror of https://github.com/okf-memory/okf-agent-memory.git synced 2026-10-01 18:25:09 +00:00

sknr 64d8fb4323 docs(knowledge): add dual-memory architecture concept and update doc source references

- Add knowledge/convention/dual-memory-architecture.md with graph links to principles and layers
- Update relative source references in knowledge concepts to point to categorized docs paths
- Record changes in knowledge/log.md and sync knowledge/index.md

2026-09-15 11:48:55 +02:00

5.1 KiB

Raw Permalink Blame History

type

title

description

resource

Why OKF Agent Memory? — Key Selling Points & Value Proposition

The OKF Agent Memory project provides a modern, standardized, and portable alternative to proprietary vector memory services and unstructured flat markdown files.

mindmap
  root((OKF Agent Memory))
    100% Git-Native & Open
      Zero Vendor Lock-in
      Plain Markdown + YAML
      Auditable via git diff
    Eliminates Memory Rot
      Progressive Disclosure
      Search Before Write
      No Context Explosion
    Trust & Provenance
      generated vs. verified
      Preserves Uncertainty
      Human Override Supremacy
    Blazing Fast Tooling
      Compiled Go Binary
      Built-in MCP Server
      Auto-Bookkeeping
    Domain-Neutral
      Coding, Coaching, Books
      Research, Operations
      Custom Taxonomies

1. Zero Vendor Lock-in & 100% Git-Native

Everything is a file: All knowledge lives in standard Markdown files with YAML frontmatter inside the repository (knowledge/).
No external database required: No Pinecone, Weaviate, or Postgres infrastructure needed to run, review, or edit memory.
Full human-in-the-loop control: Complete auditability via standard git log and git diff.

2. Built on an Open Standard (Google OKF v0.2)

Leverages the open, vendor-neutral Open Knowledge Format (OKF) v0.2 specification.
Compatible with any AI provider (Anthropic Claude, Google Gemini, OpenAI GPT, local LLMs) and IDE (Cursor, VS Code, JetBrains).

3. Solves Context Bloat & Memory Rot

Progressive Disclosure: Agents navigate via structured index.md files and relative concept links, loading only the exact context required instead of dumping megabytes of text into the prompt.
Search-Before-Write: Strictly prevents duplicated or contradictory entries by mandating that agents search existing concepts before creating new ones.

4. Built-in Trust & Provenance Layer

Explicit Trust Tiers: Clear separation between generated (written by an agent) and verified (confirmed by a human or test process).
Preserves Uncertainty: Distinguishes direct evidence from agent inference, preventing hallucinations from becoming canonical project truth.

5. High-Performance, Zero-Dependency Tooling (Go CLI & MCP)

Sub-Millisecond Execution Performance:
- In-Memory BM25 Search: Instant ranked concept retrieval in < 300 µs (microseconds).
- Full Bundle Loading & Graph Parsing: Loads 50+ concepts, builds bidirectional dependency graphs, and checks link integrity in ~4 ms.
- Single Native Binary: Zero runtime dependencies, no Python VM or Node/Deno startup overhead (instant < 5ms CLI invocations vs. 150–400ms runtime boot latency).
Native OKF v0.2 Validator: Complete native Go engine of Google's specification and graph validator, executing strict conformance checks in sub-millisecond time.
Zero API Costs for Memory Retrieval: Local lexical BM25 indexing eliminates recurring vector embedding API costs and network roundtrips.
Self-Contained Single Binary: Ships with embedded skills and templates (//go:embed), instantly runnable across macOS, Linux, and Windows via okf bootstrap.
Native MCP Server (okf mcp): Plug-and-play Model Context Protocol integration over stdio for Claude Code, Cursor, and any MCP-compliant agent client with zero daemon overhead.
Automatic Bookkeeping: Deterministically maintains log.md change logs and index.md navigation listings during write operations.

6. Truly Domain-Neutral

Designed from the ground up to support diverse workflows:
- Software Engineering: Architectural decisions, API discoveries, debugging runbooks.
- Coaching & Consulting: Client histories, session insights, goal tracking.
- Research & Writing: Literature reviews, citation tracking, reading histories.
- Operations & DevOps: Incident post-mortems, playbooks, environmental constraints.

7. Clean Separation of Cognition vs. Syntax

The LLM focuses purely on understanding, synthesizing, and reasoning.
Deterministic Go tooling guarantees syntax correctness, link integrity, and format conformance.

5.1 KiB Raw Permalink Blame History