Knowledge Consolidation Framework

When This Activates

This skill activates when:

• User wants to remember something for future sessions
• A significant decision or pattern emerges
• A bug fix reveals a gotcha
• Session learnings should be captured

The RETRIEVE→JUDGE→DISTILL→CONSOLIDATE Cycle

1. RETRIEVE

Find similar past situations from corrections/decisions:

memory_sessions category=decision "authentication pattern"
reasoning_query context="login flow architecture"

2. JUDGE

Assess if past solution applies to current context:

• Domain match? (same tech area)
• Context similarity > 30%?
• Confidence threshold met?

3. DISTILL

Extract generalizable patterns from specific instances:

• Find common trigger terms across examples
• Identify shared solution approaches
• Require 2+ examples before creating pattern

4. CONSOLIDATE

Update long-term memory with new patterns:

• Merge similar trajectories
• Update confidence scores
• Prune outdated patterns

Observation Categories

When capturing learnings, categorize them:

Category	When to Use	Example
decision	Technical choices made	"Chose native Ollama over Docker for Metal GPU"
pattern	Patterns discovered/applied	"Use host.docker.internal for Docker→host"
bugfix	Bugs found and fixed	"Fixed Firebase error by mounting key"
gotcha	Tricky/unexpected things	"Docker can't use Metal GPU on Mac"
feature	Features implemented	"Added doc_query tool to gateway"
implementation	How something was built	"Integrated AnythingLLM via REST"

Recording Workflow

When the user says "remember this" or similar:

1. Identify the Learning Type

"Remember: Docker containers can't use Metal GPU"
→ Category: gotcha
→ Domain: docker

2. Structure the Observation

{
  "category": "gotcha",
  "observation": "Docker containers on macOS cannot use Metal GPU - must use native services",
  "context": "Trying to run Ollama in Docker for GPU acceleration",
  "files": [],
  "project_id": "claude-dash"
}

3. Check for Related Past Learnings

reasoning_query context="Docker GPU Metal macOS"

4. Confirm and Store

"Recorded as a gotcha. This will surface next time Docker GPU topics come up."

Pattern Extraction

After multiple similar observations, patterns emerge:

Observations:
1. "Docker can't use Metal GPU"
2. "Ollama must run native for GPU"
3. "MLX requires native execution"

Distilled Pattern:
{
  "id": "docker_metal_gpu",
  "domain": "docker",
  "trigger_terms": ["docker", "metal", "gpu", "macos"],
  "solution_terms": ["native", "host", "not container"],
  "description": "When [docker, metal, gpu], try [native, host execution]",
  "confidence": 0.8
}

Proactive Injection

The system auto-injects relevant memories via:

• <semantic-memory> tags on prompt submission
• <past-corrections> when similar topics arise
• Pattern matching on trigger terms

Manual Queries

Check what's been learned:

memory_sessions category=decision limit=5
memory_sessions category=gotcha query="docker"
reasoning_query context="current problem description"

Consolidation Triggers

Consolidation runs:

• After 10+ new trajectories
• When explicitly requested
• During background worker runs

Output: Merged patterns, updated confidence scores, pruned stale entries.