Skip to main content
Glama
cognos-session-memory

cognos-session-memory

CognOS Session Memory

mcp-name: io.github.base76-research-lab/cognos-session-memory

Verified context injection via epistemic trust scoring for LLMs.

Solves session fragmentation by maintaining verified, high-confidence session context between conversations.

Problem

Large language models suffer from session fragmentation: each new conversation starts without verified context of previous work. This forces repeated explanations, loses decision history, and breaks long-running workflows.

Existing solutions (persistent memory systems, vector retrieval) either:

  • Lack trust scores before injection → hallucinations propagate

  • Don't audit which context was injected → compliance gaps

  • Treat all past information equally → noise overwhelms signal

Related MCP server: Memento

Solution

A plan-mode gateway that:

  1. Extracts structured context from 3-5 recent traces

  2. Scores context quality via CognOS epistemic formula: C = p · (1 − Ue − Ua)

  3. Injects as system prompt only if C > threshold

  4. Flags for manual review if C < threshold

  5. Audits every context injection with trace IDs → EU AI Act compliance

Architecture

recent_traces (n=5)
    ↓
extract_context() → ContextField + coverage
    ↓
compute_trust_score(p, ue, ua) → C, R, decision
    ↓
if C > threshold:
    system_prompt ← inject
else:
    flagged_reason ← manual review

Core Formula

C = p · (1 − Ue − Ua)
R = 1 − C

where:
  p   = prediction confidence (coverage of required fields)
  Ue  = epistemic uncertainty (divergence between traces)
  Ua  = aleatoric uncertainty (mean risk in traces)

Action Gate

R < 0.25       → PASS      (inject without review)
0.25 ≤ R < 0.60 → REFINE   (inject with caution)
R ≥ 0.60       → ESCALATE  (flag for manual review)

API

POST /v1/plan

Extract and score context.

Request:

{
  "n": 5,
  "trust_threshold": 0.75,
  "mode": "auto"
}

Response (if injected):

{
  "status": "injected",
  "trust_score": 0.82,
  "confidence": 0.82,
  "risk": 0.18,
  "decision": "PASS",
  "context": {
    "active_project": "CognOS mHC research",
    "last_decision": "Verify P1 hypothesis",
    "open_questions": ["How does routing entropy scale?"],
    "current_output": "exp_008 complete",
    "recent_models": ["gpt-4", "claude-3", "mistral"]
  },
  "system_prompt": "## CognOS Context...",
  "trace_ids": ["uuid-1", "uuid-2", ...]
}

Response (if flagged):

{
  "status": "flagged",
  "trust_score": 0.45,
  "decision": "REFINE",
  "flagged_reason": "Trust score 0.45 below threshold 0.75. Manual review recommended.",
  "trace_ids": [...]
}

Modes

  • auto (default) — inject if trust_score ≥ threshold, else flag

  • force — always inject (for testing)

  • dry_run — compute score but never inject

Claude Code Integration

As a /compact replacement

# In any Claude Code session:
/save

Claude writes a structured summary, trust-scores it, and persists it to SQLite. Next session: automatically injected as SESSION_CONTEXT before your first prompt.

See docs/COMPACT_ALTERNATIVE.md for a full comparison.

As an MCP server

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "cognos-session-memory": {
      "command": "python3",
      "args": ["/path/to/cognos-session-memory/mcp_server.py"]
    }
  }
}

Tools exposed:

Tool

Description

save_session(summary, project?)

Trust-score and persist a session summary

load_session(threshold?)

Retrieve last verified context (default threshold: 0.45)


Quick Start

Installation

git clone https://github.com/base76-research-lab/cognos-session-memory
cd cognos-session-memory
pip install -e .

Run Gateway

python3 -m uvicorn --app-dir src main:app --port 8788

Test /v1/plan (dry_run)

curl -X POST http://127.0.0.1:8788/v1/plan \
  -H 'Content-Type: application/json' \
  -d '{"n": 5, "mode": "dry_run"}'

Test /v1/plan (auto)

curl -X POST http://127.0.0.1:8788/v1/plan \
  -H 'Content-Type: application/json' \
  -d '{"n": 5, "trust_threshold": 0.75, "mode": "auto"}'

Modules

  • trust.py — CognOS confidence formula, action gate, signal extractors

  • trace_store.py — SQLite persistence (write/read/purge)

  • plan.py — Context extraction, trust scoring, system prompt building

  • main.py — FastAPI gateway + middleware

  • mcp_server.py — MCP stdio server (save_session, load_session)

Testing

pytest tests/ -v --cov=src

Documentation

Research Paper

See docs/PAPER.md — "Verified Context Injection: Epistemically Scored Session Memory for Large Language Models"

Status: Independent research — Base76 Research Lab, 2026 Authors: Björn André Wikström (Base76)

Citation

@software{wikstrom2026cognos,
  author = {Wikström, Björn André},
  title = {{CognOS Session Memory}: Verified Context Injection via Epistemic Trust Scoring},
  year = {2026},
  url = {https://github.com/base76-research-lab/cognos-session-memory}
}

License

MIT

Contact

Available Tools

2 tools
load_sessionA

Load the last verified session context from CognOS Session Memory. Returns the summary and an inject-ready SESSION_CONTEXT block. Only returns context that passed the trust threshold — never injects low-confidence or stale sessions.

Inputs: threshold.

save_sessionA

Save a trust-scored session summary to CognOS Session Memory. Call this instead of or after /compact to preserve verified context across sessions. The summary is scored epistemically — short or vague summaries get lower trust scores and may not be injected next session.

Inputs: summary (required), project.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 2 tools

    Disambiguation5/5

    The two tools have completely distinct purposes: load_session retrieves context and save_session stores it. There is no overlap in their functionality, making it easy for an agent to select the correct one.

    Naming Consistency5/5

    Both tools follow a consistent verb_noun pattern (load_session, save_session), which is clear and predictable. No mixing of conventions.

    Tool Count3/5

    With only 2 tools, the surface is thin for a memory management system. While it covers the basic load/save operations, the count is borderline low for the implied scope.

    Completeness3/5

    The tool set covers loading and saving sessions, but lacks operations like list, delete, or update sessions. There are notable gaps that may require agents to work around missing functionality.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Provides persistent, cross-session memory for AI agents, allowing them to store and automatically retrieve information across different conversations and sessions without repeating context.
      54 npm
      179
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides persistent memory for AI tools by building a local knowledge graph from conversations, enabling cross-session recall and context awareness without cloud dependencies.
      9
      MIT