HipCortex Memory Engine & Cognitive OS

Give your AI coding assistant persistent, cross-session causal memory with a full cognitive OS substrate — transactional belief revision, multi-agent workspaces, world-model rollout, DigitalTwin simulation, and topological graph tools.

VSIX 0.9.1 (stability patch) · server/pip/npm 0.9.0. 837 tests pass (339 lib + 320 unit + 128 integration + 50 property). See docs/channels.md.

Zero-config onboarding (no Rust or Cargo required)

Install from Marketplace / Open VSX / GitHub release VSIX. Extension starts a local Rust webserver under ~/.hipcortex-vscode/bin/ (or uses hipcortex.apiUrl).

• Zero external DB / Docker for default petgraph path
• Local-first storage under ~/.hipcortex-vscode/storage
• Auto-recovery: restarts server before queries when down
• Executable bundled bins: chmod 0755 applied on macOS/Linux (fixes spawn EACCES)
• Passive capture: saves code edits and terminal output automatically when hipcortex.passiveCapture is true

code --install-extension hipcortex-memory-0.9.1.vsix

What's new in v0.9.1 — Stability patch

What's new in v0.9.0 — Continuous Substrate

v0.9.0 adds a continuous dynamical simulation layer on top of the v0.8.0 Cognitive OS substrate.

DigitalTwin (RK4 + HybridRollout)

• POST /v1/twin — create a DigitalTwin fork with configurable state dimension, dt, and max covariance
• POST /v1/twin/:id/step — advance one step via RK4 integrator; returns continuous state vector
• POST /v1/twin/:id/rollout — multi-action HybridRollout; returns continuous_trajectory + continuous_sigma_norm
• GET /v1/twin/:id — inspect trajectory depth and record count
• VS Code commands: Create DigitalTwin, DigitalTwin: Step, DigitalTwin: Rollout, DigitalTwin: Show State

ExperienceStore (3-tier pyramid)

• Raw tier — unprocessed Temporal records
• Episode tier — Skill/Belief records with evidence links
• Abstract tier — consolidated Temporal records (action="consolidated")
• AutoConsolidate achieves ≥ 90% hot-set reduction while preserving full provenance
• GET /v1/experience/:actor/tiers — tier counts + compression ratio + raw pressure flag
• POST /v1/experience/:actor/search — semantic search across all tiers
• VS Code command: Show Experience Tier Stats

Python SDK: HipCortexSubstrate

from hipcortex import HipCortexSubstrate
sub = HipCortexSubstrate("http://localhost:3030")
twin_id = sub.create_twin(dim=4, dt=0.1)
state   = sub.twin_step(twin_id, "move_forward")
tiers   = sub.experience_tiers("my-agent")

MCP: 5 new tools + 4 new resources (42 total / 7 resources)

• twin_create, twin_step, twin_rollout, twin_get, experience_tiers
• Resource: hipcortex://experience/tiers auto-injected at session start

What's new in v0.8.0 — Cognitive OS Substrate

v0.8.0 reframes HipCortex from a memory store into a closed dynamical cognitive substrate. All state evolution flows through a single transactional operator family (CognitiveDelta).

Transactional Core

• POST /v1/cognitive/transact — atomic delta application with CoherenceChecker + TxLog
• GET /v1/cognitive/snapshot — typed snapshot with tx_cursor, beliefs, goals
• POST /v1/state/diff — causal ΔS between any two tx indices

JTMS Belief Revision

• Doyle-style justification-based truth maintenance
• RetractBelief / AssertJustification — cascade retraction through dependency graph
• No belief stays IN without a valid justification

Causal Motif Compactor

• Frequent causal path mining → SkillPayload + high-confidence Belief induction
• AutoConsolidate { min_frequency } — sub-linear memory growth on long trajectories

Multi-Agent Workspaces

• WorkspaceOpen { mode: Private | Shared } + WorkspaceMerge deltas
• OR-Set CRDT merge for concurrent mutations; no silent contamination

Rollout + Drift Alarms

• Multi-step Kalman covariance expansion (discrete Lyapunov recursion)
• Continuous goal-distance drift alarm on POST /v1/fork/:id/rollout

@hipcortex chat commands

Open Copilot / Antigravity chat and type @hipcortex:

• @hipcortex health — server status, calibration score, epistemic entropy
• @hipcortex add <content> — store decision / preference / constraint
• @hipcortex query <query> — semantic + topological retrieval
• @hipcortex status — Headroom vs Caveman mode and savings

Language Model Tools (10)

Extension registers 10 tools with vscode.lm (requires host LM tool API):

VS Code Commands (15)

Phase-5 Operator Methods (8)

Available as HipCortexClient TypeScript methods and wired to VS Code commands:

MCP Integration (42 tools, 7 resources)

MCP hosts (Claude Code, Cursor, Windsurf, …) use the Python MCP server via hipcortex install.
42 tools + 7 auto-injected resources:

• hipcortex://context/relevant — top-k semantically relevant memories
• hipcortex://beliefs/current — active belief records
• hipcortex://context/conversation — recent temporal traces
• hipcortex://experience/tiers — NEW ExperienceStore tier stats for current actor

Register in .mcp.json:

{
  "mcpServers": {
    "hipcortex": {
      "type": "stdio",
      "command": "python",
      "args": ["/path/to/hipcortex/sdk/mcp/server.py"],
      "env": { "HIPCORTEX_URL": "http://localhost:3030" }
    }
  }
}

Headroom & Caveman (token savings)

• Headroom (Top-5): ~59–84% token reduction vs full history dump
• Caveman (Top-3): ~70–88% in tight debug loops

Configuration (settings.json)

{
  "hipcortex.apiUrl": "http://127.0.0.1:3030",
  "hipcortex.apiKey": "",
  "hipcortex.autoStart": true,
  "hipcortex.optimizationMode": "headroom",
  "hipcortex.passiveCapture": true
}

Local development & packaging

cd vscode-extension
npm install
npm run compile
npm test
npx @vscode/vsce package --no-dependencies

Produces hipcortex-memory-0.9.1.vsix (version from package.json).

Related

• Channel honesty: docs/channels.md · hipcortex channels
• Capability matrix: docs/capabilities.md
• Host wizards: docs/hosts/README.md
• Architecture: docs/architecture.md