FEATURES
Three-signal retrieval. Local embeddings. Wiki-link knowledge graph. MCP-native. No cloud, no API keys, no vendor lock-in.
Most memory tools use one search method. Ori combines three — and the combination is what makes it work.
Semantic similarity via local embeddings (Xenova/all-MiniLM-L6-v2). Finds conceptually related notes even when the vocabulary differs. "Token incentives" matches "engagement mechanics." Runs entirely on your machine — no API, no latency, no cost.
Traditional full-text search for exact terms and phrases. Fast, deterministic, catches things semantic search misses. Function names, specific error messages, exact quotes — keyword search finds them reliably when you need precision over fuzzy recall.
Wiki-links between notes create a knowledge graph. When a search hits a note, activation spreads to connected notes through those links — surfacing context that neither keyword nor vector search would find alone. The whole is greater than the sum of its parts.
Every competitor stores your memory in an opaque database. Ori stores it as plain markdown files you can read, edit, and own forever.
---
description: Redis gives 10x lower latency
than Postgres for session state
type: decision
project: [backend]
status: active
created: 2026-03-04
---
Chose Redis over Postgres for session caching
after benchmarking. Postgres p99 was 80ms,
Redis p99 was 8ms at our traffic level.
Relevant Notes:
- [[session architecture overview]]
- [[database selection criteria]]Works with Claude Code, Cursor, Windsurf, Cline, or any MCP-compatible client. Add to your config and your agent wakes up with memory.
ori_orient Session briefing — daily status, reminders, vault health, active goals. Call at session start.
ori_query_ranked Full 3-signal retrieval with intent classification and spreading activation. The main search tool.
ori_add Capture a new insight to inbox. Fast, atomic, with prose-as-title convention.
ori_promote Promote inbox note to the graph with classification, linking, and area assignment.
ori_query_similar Pure vector search — faster single-signal retrieval when you need speed over depth.
ori_query_important Notes ranked by PageRank — find the most connected, structurally important memories.
ori_update Update agent identity, goals, methodology, daily log, or reminders.
ori_health Full vault diagnostic — orphans, dangling links, schema violations, throughput.
The alternatives are well-funded. But they're building databases. We're building files.
| Ori Mnemos | Mem0 | Zep | Letta | |
|---|---|---|---|---|
| Storage | Markdown files | Cloud database | Neo4j graph DB | Proprietary |
| Local-first | ✓ Always | Optional add-on | ✗ | ✗ |
| API keys needed | None | Required | Required | Required |
| Human-readable | ✓ Plain text | ✗ | ✗ | ✗ |
| Git-versionable | ✓ Native | ✗ | ✗ | ✗ |
| MCP-native | ✓ Built for it | Bolted on | ✗ | ✗ |
| Open source | ✓ Apache-2.0 | Partial | Partial | ✓ |
| Vendor lock-in | None | High | High | Medium |
npm i -g ori-memoryori initori serve --mcp