Skip to content

Why.md

Latest commit

 

History

History
161 lines (117 loc) · 7.59 KB

Why.md

File metadata and controls

161 lines (117 loc) · 7.59 KB

Why OpenMemory

OpenMemory exists because AI memory today is fragmented, opaque, and inefficient.
Most tools either store plain embeddings without context (vector DBs) or lock memory behind closed APIs (Supermemory, OpenAI Memory).
OpenMemory solves this by building a structured, explainable, multi-sector cognitive memory engine — open source, local-first, and framework-free.

1. Why Not Just Use a Vector Database?

Vector databases like Chroma, Weaviate, or Pinecone are excellent for generic semantic search, but they fall short when used as long-term memory systems for agents or assistants.

Limitation Vector DBs OpenMemory
Contextual structure Flat embeddings only Multi-sector (episodic, semantic, procedural, emotional, reflective)
Biological alignment None Inspired by human brain’s sectorial memory formation
Graph relationships Optional (manual edges) Automatic, single-waypoint graph
Temporal awareness Manual timestamps Built-in recency & salience scoring
Explainable recall ❌ ✅ Full trace path via waypoint graph
Cost control Scales with vector count Light, SQLite-based — predictable cost
Agent integration Manual Native /memory/add + /memory/query pipeline
Privacy Cloud or managed service 100% local and auditable

In short: Vector DBs store “what was said.”
OpenMemory remembers “what it meant, when, how it felt, and why it matters.”

2. Why Not Use Supermemory or OpenAI Memory?

Factor Supermemory (SaaS) OpenAI Memory OpenMemory
Ownership Significantly closed Closed SaaS Open-source Apache 2.0
Hosting Cloud only Cloud only Self-hosted or cloud
Explainability Black-box Black-box Transparent
Architecture Flat embeddings Proprietary HMD v2 (multi-sector + single-waypoint graph)
Response latency ~350-400 ms ~300 ms 110-130 ms
Cost per 1M tokens ~$2.50+ ~$3.00+ ~$0.30-0.40
Local embeddings ❌ ❌ ✅ E5, BGE, Ollama
Integration Web API only GPT-native REST / JS / Python SDK
Focus Chatbot SaaS memory Assistant-level Developer memory infrastructure

Summary:

  • OpenMemory runs locally, for free, and integrates directly with your stack.
  • It’s 5–10× cheaper and 2–3× faster.
  • You can see how every memory is formed, stored, decayed, and recalled.

3. The Architectural Difference

Vector DBs:

  • Treat every entry as an independent embedding.
  • Retrieval = cosine similarity search.
  • No understanding of relationships, meaning, or salience.
  • Result: duplicate data, poor recall, high storage cost.

OpenMemory:

  • Breaks information into memory sectors (like the brain).
  • Stores one unified node per memory with multi-sector embeddings.
  • Adds a single, strongest waypoint between related memories.
  • Retrieval uses composite similarity and activation spreading.
  • Result: denser understanding, faster lookup, lower redundancy.

4. Cost and Efficiency

  • Vector DBs: Cost grows linearly with embeddings; storage and API calls dominate.
  • SaaS Memories: Add markup for hosting, tokens, and proprietary APIs.
  • OpenMemory: Uses SQLite + FAISS/Chroma locally.
    • 100k memories ≈ 1.5 GB
    • 1M memories ≈ 15 GB
    • Runs fine on a $5/month VPS.
Feature Vector DB SaaS Memory OpenMemory
Query time (100k) 160-250 ms 300-400 ms 110-130 ms
Cost (1M tokens w/ hosted embeddings) ~$1.20 ~$3.00+ ~$0.30-0.40
Self-hosted Partial No Yes
Local embeddings Optional No Yes

5. Cognitive Realism and Agent Readiness

OpenMemory introduces sector-specific cognition:

  • Episodic: Event memories — when/what happened
  • Semantic: Facts and preferences
  • Procedural: Habits and workflows
  • Emotional: Feelings and tone
  • Reflective: Meta-memory and logs

This structure mirrors human cognition and allows for contextual recall in LLM agents.

Example:

“User said they enjoy coding at night and feel productive.”
→ Stored across semantic (“coding preference”), emotional (“feel productive”), and episodic (“time: night”) sectors.
→ Recalled as one linked thought with a waypoint reference.

No vector DB can do this natively.

6. Developer Value Proposition

Area Why OpenMemory Wins
Setup Single binary or Docker; no external service required
Integrations Works with any LLM (OpenAI, Gemini, AWS, Ollama, Claude)
SDKs TypeScript + Python included
Cost 10× cheaper than hosted alternatives
Explainability Memory formation and recall fully transparent
Performance p95 retrieval under 130ms on 100k+ nodes
Scalability Horizontally shardable by sector
Extensibility Add new embedding models with 1 config line
Privacy Zero vendor lock-in; full data control

7. Philosophy

Memory is not a database.

It’s a dynamic system that evolves, decays, and recalls — contextually and semantically.

OpenMemory treats every memory as a living object with:

  • Time decay
  • Reinforcement (based on recall frequency)
  • Emotional intensity
  • Reflective self-linking

This turns static storage into a cognitive substrate for AI.

8. Summary

Metric OpenMemory Others
Speed 2–3× faster
Cost 6–10× cheaper
Explainability 100% transparent
Integration Any framework
Data ownership 100% yours
AI readiness Agent-first design

OpenMemory = Cognitive storage for AI agents. Not just another vector DB — it’s the brain behind your AI.

9. Caveats & Accuracy

Note: the figures and comparisons above are approximate and may not be accurate for every environment or over time.