We believe the best way to advance AI engineering is to share what we learn. Our open source work is contributed back to the communities that make this work possible.
Persistent memory for any AI — zero token cost until recalled, 79% smaller responses with TOON format
View Project →ai-memory gives any AI assistant persistent memory across sessions. Unlike built-in memory systems (Claude auto-memory, ChatGPT memory) that load your entire memory into every conversation — burning tokens and money on every message — ai-memory uses zero context tokens until recalled. Only relevant memories come back, ranked by a 6-factor scoring algorithm.
A single Rust binary with three interfaces (MCP, HTTP, CLI), four feature tiers from zero-dependency keyword search to autonomous recall with local LLMs via Ollama, and TOON format (Token-Oriented Object Notation) that cuts response tokens by 79%. Works with Claude, ChatGPT, Grok, Llama, Cursor, Windsurf, Continue.dev, and any MCP-compatible platform.
The MCP server provides recall-first prompts that teach AI clients to use memory proactively — recalling at session start, storing corrections as permanent knowledge, and using TOON compact format automatically. 158 tests across 14/14 modules with 95%+ coverage.
Every AI platform that offers built-in memory charges you tokens for it — whether you use it or not. ai-memory eliminates that tax entirely.
Built-in systems inject 200+ lines of memory into every conversation's system prompt. ai-memory uses zero context tokens until the AI explicitly calls memory_recall. You only pay for what you use.
When memories are recalled, TOON (Token-Oriented Object Notation) declares field names once as a header, then lists values as pipe-delimited rows. 3 memories: JSON = 1,600 bytes, TOON compact = 336 bytes.
FTS relevance, priority weight, access frequency, confidence, tier boost, and recency decay. Only the most relevant memories surface — no wasted tokens on irrelevant context.
Two MCP prompts (recall-first + memory-workflow) teach AI clients to use memory proactively at connection time. No manual configuration needed — the behavior is built in.
Scale from zero-dependency keyword search to autonomous memory management with local LLMs. Each tier builds on the one below it.
SQLite FTS5 full-text search. Zero ML dependencies, zero memory overhead. 13 MCP tools. The binary is entirely self-contained.
Adds dense vector embeddings (all-MiniLM-L6-v2, 384-dim) with HNSW index. Hybrid recall blends FTS5 + cosine similarity. 14 MCP tools. ~256 MB RAM.
Upgrades to nomic-embed-text (768-dim) via Ollama. Adds Gemma 4 E2B LLM for query expansion, auto-tagging, and contradiction detection. 17 MCP tools. ~1 GB RAM.
Gemma 4 E4B + neural cross-encoder reranker (ms-marco-MiniLM). Full autonomous memory reflection and contradiction resolution. 17 MCP tools. ~4 GB RAM.
autoMemoryEnabled in settings.json
codex mcp add CLI shortcut
gemini mcp add CLI shortcut
Graph-native AI skill database for the OpenClaw community — powered by Neo4j
View Project →OpenClaw Graph is AlphaOne's contribution to the open source OpenClaw project — a graph database layer that fundamentally transforms how AI agents discover, relate, and invoke skills. By replacing flat skill registries with a graph-native architecture, agents gain multi-hop reasoning across skill relationships, cluster-aware discovery, and dramatically reduced token consumption.
v1.5 introduces a Rust event-driven sync daemon that materializes Neo4j graph data into OpenClaw workspace files with FSEvents file watching for sub-second write-back. The Rust binary replaces a Python polling script — eliminating interpreter startup overhead for a 111x per-cycle speedup and 12x memory reduction. Persistent Bolt connection, 3.8 MB binary, ~4 MB RSS.
v1.5 also ships a self-contained installer — seed.py reads all 315 skills directly from the repo's SKILL.md files, requiring zero external data files. One command: ./install.sh seeds Neo4j, deploys workspace stubs, and optionally builds the Rust daemon.
The entire skill graph runs on Neo4j with proper graph modeling — SkillCluster nodes, typed relationships, namespaced labels for multi-graph coexistence, and workspace-scoped isolation. ~10 MB footprint with sub-millisecond query times.
Traditional AI skill registries are flat lists — the agent must scan everything to find what it needs, consuming tokens proportional to the registry size. Graph databases change the fundamental economics. With Neo4j integration, skills become first-class graph citizens alongside other knowledge domains.
Skills are connected through typed edges — RELATED_TO for dependencies, IN_CLUSTER for semantic grouping. Neo4j's native graph storage means traversals are index-free adjacency lookups — constant time regardless of graph size.
27 SkillCluster nodes group related skills into semantic domains. Each cluster is a proper graph node with IN_CLUSTER edges — agents retrieve coherent, pre-organized context windows instead of arbitrary skill fragments.
Namespaced labels (OCAgent, OCMemory, OCTool) let the skill graph coexist with other graph domains in a shared Neo4j instance. One database, multiple knowledge domains, workspace-level isolation.
Cypher queries naturally enforce scope boundaries via workspace properties. Agents retrieve only the subgraph relevant to their identity — no over-fetching, no token waste. Context assembly, not context accumulation.
The 229% token cost reduction comes from multiple optimization layers working together — what we call the TOON (Token-Optimized Operation Network) approach.
Graph queries return precisely scoped subgraphs instead of bulk skill dumps. Agents receive only the context they need for the current task — eliminating the token overhead of transmitting entire registries.
Graph edges encode relationships that would otherwise require verbose natural language descriptions. A single edge replaces paragraphs of context about how skills relate — compressing the context window without losing information.
Frequently accessed subgraphs are pre-materialized. Common agent workflows hit warm caches instead of executing fresh graph queries — reducing both latency and the token cost of re-computing context.
Instead of loading all skill details upfront, the graph serves summaries first and full definitions on demand. Agents decide what to expand based on initial traversal — paying token costs only for what they actually use.
The workspace materializer is a long-running Rust binary (neo4j-sync) that bridges Neo4j and OpenClaw flat files. It maintains a persistent Bolt connection and uses macOS FSEvents for instant write-back — replacing a Python polling script that spent 97% of each invocation on interpreter startup.
A single Neo4j Bolt connection stays open for the lifetime of the daemon. No per-cycle startup cost, no connection churn — 60 new connections/hour reduced to 1.
FSEvents file watching with 500ms debounce detects IDENTITY.md changes and syncs to Neo4j in under a second. Agents see identity updates reflected instantly instead of waiting up to 60 seconds.
Compiled with LTO and size optimization. The entire daemon — Neo4j driver, file watcher, async runtime, regex engine — fits in a 3.8 MB arm64 binary with ~4 MB RSS at runtime.
Drop-in replacement for the Python script. Same launchd service name, same log paths, same workspace files. Atomic swap: unload Python, load Rust — under 1 second of downtime.
One command installs everything. seed.py reads skills directly from the repo's 315 SKILL.md files and 217 relationship pairs — no external data files, no migration scripts, no manual steps.
Skills are parsed from YAML frontmatter in the repo's skills/*/SKILL.md files. Relationships loaded from seed-data/skill_rels.json. Workspace defaults embedded in Python. Only Neo4j + Python neo4j driver required.
315 skills, 27 clusters, 217 relationships, and 50 workspace nodes seeded in under 1 second. UNWIND batches of 50, MERGE for full idempotency — safe to re-run anytime.
./install.sh --workspace myagent creates isolated workspace nodes with auto-deployed stubs. Run multiple agents on a single Neo4j instance with workspace-scoped isolation.
seed.py --verify validates all node counts, relationship edges, and workspace stub queries. --dry-run previews every operation before touching Neo4j.
We're always looking for contributors and partners who share our commitment to building better AI infrastructure in the open.
sales@alpha-one.mobi