Your AI is ephemeral. Your context shouldn't rot.
Context engineering infrastructure implementing Anthropic's published principles for building capable agents.
AI agents generate massive thinking data. Without management, old thoughts compete with new conclusions for the attention budget—a phenomenon called context rot. Maenifold prevents this through:
- Just-in-time retrieval:
[[WikiLinks]]as lightweight identifiers, not payloads - Compaction: Memory cycles consolidate episodic → semantic
- Decay: Time-based signal-to-noise management (calibrated to iteration speed)
- Structured note-taking: Sequential thinking with graph integration
- Sub-agent architectures: Clean contexts, condensed handoffs
Six composable layers: [[WikiLinks]] → Graph → Hybrid Search → Session State → Reasoning → Orchestration. No magic. No fake AI. Space for intelligence to emerge.
Install
# Homebrew (macOS/Linux) — handles PATH automatically
brew install msbrettorg/tap/maenifold
# Manual — download from GitHub Releases, extract, add to PATH or symlink
# https://github.com/msbrettorg/maenifold/releases/latestRequirements
| Platform | Binary | Notes |
|---|---|---|
| macOS | osx-arm64, osx-x64 | Apple Silicon or Intel; Homebrew recommended |
| Linux | linux-x64, linux-arm64 | x64 or ARM64 |
| Windows | win-x64 | x64 only; use PowerShell syntax below |
Binaries are self-contained (.NET 9.0 bundled). Vector embeddings use ONNX runtime (bundled). No external dependencies.
Use the CLI
macOS/Linux (bash):
maenifold --tool WriteMemory --payload '{"title":"Auth Decision","content":"Using [[OAuth2]] for [[authentication]]"}'
maenifold --tool SearchMemories --payload '{"query":"authentication","mode":"Hybrid"}'
maenifold --tool BuildContext --payload '{"conceptName":"authentication","depth":2}'
maenifold --tool Sync --payload '{}'Windows (PowerShell):
maenifold --tool WriteMemory --payload '{\"title\":\"Auth Decision\",\"content\":\"Using [[OAuth2]] for [[authentication]]\"}'
maenifold --tool SearchMemories --payload '{\"query\":\"authentication\",\"mode\":\"Hybrid\"}'
maenifold --tool BuildContext --payload '{\"conceptName\":\"authentication\",\"depth\":2}'
maenifold --tool Sync --payload '{}'The CLI is the primary interface. See SCRIPTING.md for advanced patterns.
Build domain expertise — See Bootstrap Guide for the full journey: research workflows, custom roles, custom workflows, and maintenance.
MCP vs CLI
Both interfaces have full feature parity. Choose based on your use case:
| Interface | Advantage | Best for |
|---|---|---|
| MCP | Auto-sync watcher keeps graph current | Interactive sessions, simple queries |
| CLI | Filter intermediate results, preserve context (why this matters) | Complex workflows, scripting, Graph-RAG patterns |
MCP config (for Claude Code, Claude Desktop, etc.):
{
"mcpServers": {
"maenifold": { "command": "maenifold", "args": ["--mcp"], "type": "stdio" }
}
}| Skill | What You Get |
|---|---|
| Maenifold | 25+ tools, 6-layer composable layers, sequential thinking, 35+ workflows |
| Product Manager | maenifold-aware agents, graph context injection, quality gates, sprint traceability |
Layered from theory → memory → symbols → reasoning → orchestration → assets
| Concept | Origin | Application |
|---|---|---|
| Ma (間) | Japanese aesthetics | The space between things as the thing itself (what we don't do) |
| Concept | Origin | Application |
|---|---|---|
| Context Engineering | Anthropic (2025) | Attention budget management: just-in-time retrieval, compaction, decay, structured notes, sub-agents |
| ACT-R | Anderson (CMU); Wixted & Ebbesen (1991) | Power-law inspired decay (exponential approximation) |
| New Theory of Disuse | Bjork & Bjork | Storage vs retrieval strength |
| Two-Stage Memory | Cognitive neuroscience | Episodic → semantic consolidation |
| Linguistic Relativity | Weak form (Boroditsky, 2003) | Perspectives change the linguistic frame of LLM reasoning |
| ConfessionReport | OpenAI (Barak et al., 2025) | Inference-time honesty enforcement via hooks + adversarial audit |
| Feature | What It Means | Implementation |
|---|---|---|
| Two-Stage Model | Fast episodic encoding → slow semantic consolidation | memory://thinking/ (episodic) vs memory://research/ (semantic) |
| ACT-R Decay | Memories fade without access following power-law | DecayCalculator: base × time^(-0.5) |
| Storage vs Retrieval | Pointers persist; only accessibility fades | WikiLinks never deleted; decay affects ranking |
| Maintenance Cycles | Periodic graph hygiene | 4 workflows: consolidation, decay, repair, epistemic |
| Consolidation | Thinking sessions → semantic memory | Replay via RecentActivity, distill to WriteMemory, link via FindSimilarConcepts |
| Feature | What It Means | Implementation |
|---|---|---|
| Concept-as-Protocol | WikiLinks carry context between agents | PreToolUse hook extracts [[concepts]], injects graph context |
| Lazy Graph | No predefined schema; structure emerges | Sync extracts WikiLinks; co-occurrence creates edges |
| Hybrid Search | Semantic similarity + exact matching | SearchMemories fuses with Reciprocal Rank Fusion (k=60) |
| Concept Repair | Normalize WikiLink variants safely | RepairConcepts validates similarity ≥0.7 before replacing |
| Feature | What It Means | Implementation |
|---|---|---|
| Sequential Thinking | Multi-step reasoning with persistence, graph integration, multi-agent branching | Extends MCP server with sessionId, branchId, parentWorkflowId |
| Assumption Ledger | Track beliefs and their validation status | AssumptionLedger with confidence levels, evidence links |
| Feature | What It Means | Implementation |
|---|---|---|
| Workflow Engine | Structured multi-step methodology execution | Workflow tool with JSON steps, toolHints, guardrails |
| Nested Composition | Workflows invoke other workflows and tools | Workflows embed SequentialThinking; bidirectional linking |
| Session Persistence | Resume reasoning across days | State in memory://workflow/; session IDs enable continuation |
Notable Workflows:
| Workflow | What It Demonstrates | How |
|---|---|---|
| Workflow Dispatch | Meta-cognitive methodology selection | Analyzes problem characteristics, queries graph for similar past problems, selects optimal workflow |
| Research Think Tank | Multi-agent collaborative reasoning | Defines roles (synthesizer, critic, explorer), structures knowledge construction phases |
| Agentic SLC | Quality-controlled development | Embeds anti-slop checks, requires RTM traceability, enforces ConfessionReports |
| Constitutional Roles | Constitutional AI for persona creation | Uses principles + examples to generate role definitions that constrain agent behavior |
| Higher-Order Thinking | Meta-cognitive reflection | Recursive self-analysis steps: examine reasoning, identify biases, refine approach |
| Six Thinking Hats | Structured perspective switching | Sequences through DeBono's colors with explicit transitions and synthesis |
Multi-agent orchestration requires an MCP client with agent dispatch (Claude Code, Codex, aishell) or CLI scripting with subprocess spawning.
| Asset | Count | Examples |
|---|---|---|
| Workflows | 35+ | Deductive, design thinking, agentic sprints, game theory |
| Roles | 16 | Architect, PM, red-team, blue-team, researcher |
| Thinking Colors | 7 | DeBono's Six Hats + Gray (skeptical inquiry) |
| Perspectives | 12 | Native language modes for culturally-aware reasoning |
Claude Desktop — Same MCP config, different location:
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json
Custom data directory:
export MAENIFOLD_ROOT=~/my-knowledge-base| Spec | Value |
|---|---|
| Language | C# / .NET 9.0 |
| Vectors | 384-dim (all-MiniLM-L6-v2 via ONNX) |
| Search | Reciprocal Rank Fusion (k=60) |
| Database | SQLite + sqlite-vec (bundled) |
| Memory Cycle | 24h compaction interval; decay params expressed as cycle multiples (volume-dependent, scriptable) |
| Decay Model | ACT-R power-law (d=0.5); calibrated to memory cycle: sequential (2-3×), workflows (7×), memory (14-28×) |
| Memory Tiers | Grace periods anchored to iteration speed: Sequential (2-3d) / Workflows (7d) / Semantic (14-28d) |
| Maintenance | 4 workflows (consolidation/decay/repair/epistemic) mirror biological sleep phases |
| Scale | > 1M relationships tested |
Decay affects search ranking only. Files are never deleted.
Local-only, single-user, no authentication. Memory files (markdown) are the source of truth; the SQLite database is a regenerable cache (maenifold --tool Sync). Version control memory with git to track changes alongside code. Set MAENIFOLD_ROOT to store memory in your repository for atomic commits of code + reasoning.
- Bootstrap Guide — From empty graph to domain expertise
- Scripting Guide — CLI patterns, Graph-RAG, HYDE, FLARE
- Complete Guide — Architecture, philosophy, detailed examples
- Maenifold Skill — Core tools and composable layers
- Product Manager Skill — Multi-agent orchestration
MIT License. Contributions welcome at github.com/msbrettorg/maenifold.
