Memory Enhancement

Manage citations, verify code references, and track confidence scores for Serena memories. Ensures memories stay accurate by linking them to specific code locations and detecting when those locations change.

Triggers

• add citation to memory - Link memory to specific code location
• verify memory citations - Check if code references are still valid
• check memory health - Generate staleness report across all memories
• update memory confidence - Recalculate trust score based on verification

Quick Reference

Input	Output	Duration
Memory ID + code reference	Citation added with validation	< 5 seconds
Memory directory	Health report with stale memories	< 30 seconds
Verification results	Updated confidence scores	< 10 seconds

Decision Tree

Need memory enhancement?
│
├─ Add citation to memory → edit the memory file (no CLI command)
├─ Verify citations → verify or verify-all command
├─ Check memory health → health command
├─ Traverse memory graph → graph command
└─ Show confidence scores → confidence command

Command Surface

Global options come before the subcommand. Placing --repo-root or --memories-dir after the subcommand exits 2.

python -m memory_enhancement [--repo-root PATH] [--memories-dir PATH] <command>

Command	Options	Exit codes
verify	--memory-id ID	1 if any citation is invalid, or the id is unknown
verify-all	--json	1 if any citation is invalid
health	--json, --text, --markdown (default)	1 if any citation is broken or stale, or any memory is stale
graph	--start ID (required), --depth N	1 if the start id is unknown
confidence	none	0
search	QUERY (positional), --top N, --json	0

--memories-dir must resolve inside --repo-root; otherwise the CLI exits 1. A memory id is its path under the memories directory without the .md suffix, so .serena/memories/testing/foo.md has the id testing/foo.

Process

Phase 1: Identify Target Memory

Locate the memory file by id:

1. Derive the id - The id is the path under the memories directory without the .md suffix (.serena/memories/testing/foo.md has id testing/foo).
2. Point at a different tree - Pass --memories-dir PATH before the subcommand. It must resolve inside --repo-root.
3. Validate existence - verify --memory-id <id> prints Memory not found: <id> and exits 1 when the id is unknown.

Verification: Memory file exists and is readable

Phase 2: Add/Verify Citations

Add Citation

There is no add-citation command. Citations are markdown, so add one by editing the memory file and appending a citation line to its body:

[cite:file](src/api.py) - the error handler lives here

Syntax: [cite:<source-type>](<target>) - <context>

• source-type - one of file, function, issue, pr, adr, memory, url
• target - repository-relative path or identifier; must be non-empty
• context - optional prose after -, kept as the citation context

An unrecognized source-type warns on stderr and the citation is dropped, so the memory can still exit 0 with the citation missing. Verify after editing. A citations: block in YAML frontmatter is not read by this tool.

Verification: verify --memory-id <id> lists the new citation

Verify Citations

# Single memory
python -m memory_enhancement verify --memory-id <memory-id>

# All memories, human-readable
python -m memory_enhancement verify-all

# All memories, JSON for CI
python -m memory_enhancement verify-all --json

Output Indicators:

• [PASS] <target> - <reason> - citation resolves
• [FAIL] <target> - <reason> - citation is broken or stale
• Exit code 1 when any citation is invalid

Verification: Citations validated against current codebase state

Phase 3: Show Confidence

Recalculate scores from current verification results:

python -m memory_enhancement confidence

Prints <memory-id>: <score> for every memory. The command computes scores on each run and does not write them back to the memory files, and it has no single-memory filter.

Confidence Calculation:

A weighted blend of four factors, clamped to 0.0-1.0:

Factor	Weight	Definition
Citation validity	0.50	valid citations / total citations; 1.0 when there are none
Update recency	0.25	decays linearly to 0 over 90 days since updated_at
Link count	0.15	outgoing links / 10, capped at 1.0
Memory freshness	0.10	decays linearly to 0 over 365 days since created_at

A memory with no citations therefore does not score low; it scores high on validity and is limited only by age and link count. Read a high score as "nothing is known to be broken", not "this was checked against code".

Interpretation:

Score Range	Meaning	Action
0.9 - 1.0	High confidence	Trust memory, use in decisions
0.7 - 0.9	Medium confidence	Review stale citations
0.5 - 0.7	Low confidence	Update memory or mark obsolete
0.0 - 0.5	Very low confidence	Memory likely outdated

Verification: confidence prints a score for every memory

Phase 4: Report Results

Display summary with actionable recommendations:

List Citations

There is no list-citations command. Verifying a memory prints its citations alongside their status:

python -m memory_enhancement verify --memory-id <memory-id>

Human-readable output:

testing/demo:
  [PASS] src/api.py - File exists
  [FAIL] src/missing.py - File not found: src/missing.py

For machine-readable output across every memory, use verify-all --json. It emits a flat array, one object per citation:

[
  {
    "memory_id": "testing/demo",
    "target": "src/api.py",
    "source_type": "file",
    "valid": true,
    "reason": "File exists"
  }
]

Health Report

python -m memory_enhancement health --json

Reports the following, and nothing else:

• total_memories, total_citations
• valid_citations, stale_citations, broken_citations, unverified_citations
• health_score ((valid + 0.5 * stale) / total; 1.0 for an empty corpus, 0.0 when memories exist but carry no citations)
• stale_memories (ids only, unordered)
• recommendations (prose strings)

There is no orphan detection and no staleness ranking.

Verification: Report generated successfully

Script Reference

Operation	CLI Command	Key Parameters
Add citation	none; edit the memory body	[cite:<type>](<target>) - <context>
Verify memory	python -m memory_enhancement verify	--memory-id
Verify all	python -m memory_enhancement verify-all	--json
Health report	python -m memory_enhancement health	--json, --text, --markdown
Show confidence	python -m memory_enhancement confidence	none
List citations	python -m memory_enhancement verify	--memory-id
Graph traversal	python -m memory_enhancement graph	--start, --depth
Search memories	python -m memory_enhancement search	QUERY, --top, --json

Prefix any of these with --repo-root PATH or --memories-dir PATH to point at a different tree. Both are global options and must precede the subcommand.

Anti-Patterns

Avoid	Why	Instead
Placing --repo-root after the subcommand	argparse rejects it and exits 2	Put global options first
Adding a citation without checking the target	Nothing validates on write; it is a text edit	Run verify --memory-id <id> right after editing
Putting citations in YAML frontmatter	The parser only reads [cite:...] in the body	Add the citation line to the memory body
Reading a high confidence score as verified	Memories with no citations score high on validity	Check total_citations before trusting a score
Using absolute paths	Breaks on different machines	Use repo-relative paths
Forgetting to verify after refactoring	Citations go stale silently	Run verify-all regularly or in CI

Integration with Existing Skills

• reflect - Auto-capture citations from learnings that reference code
• memory - Verify citations during memory search
• curating-memories - Update citations when memories change
• qa - Run verification as part of test strategy

Phase 5: Health Reporting

Run batch health checks:

# Full report (human-readable markdown, the default)
python -m memory_enhancement --repo-root . --memories-dir .serena/memories health

# JSON output (for CI parsing)
python -m memory_enhancement health --json

# Markdown output (for PR comments)
python -m memory_enhancement health --markdown

# Plain text output
python -m memory_enhancement health --text

The three format flags are mutually exclusive. There is no exemption mechanism: exempt: true in frontmatter has no effect, and every memory in the directory is scanned.

Exit Codes:

• 0: No broken citations, no stale citations, no stale memories
• 1: One or more memories are broken or stale
• 2: argparse rejected the command line

CI Integration

Memory Health Workflow

The .github/workflows/memory-health.yml workflow runs health checks on all PRs:

• Detects changes to .serena/memories/** and memory enhancement code
• Generates JSON and Markdown health reports
• Posts/updates a PR comment with results
• Non-blocking (warning only, not a required check)
• Uses <!-- MEMORY-HEALTH --> marker for idempotent comment updates

Citation Verification Workflow

The .github/workflows/citation-verify.yml workflow verifies citations:

• Runs on every pull request to main, plus workflow_dispatch
• Filters on .serena/memories/** and .claude/skills/memory-enhancement/**
• Runs python3 -m memory_enhancement --repo-root . --memories-dir .serena/memories verify-all
• Blocking: a stale or broken citation exits 1 and fails the check
• Posts no PR comment
• Pairs with a skip-verification job so the required check still reports on pull requests that touch neither path

Verification

After using this skill:

• Citations validated against the current codebase
• Confidence scores read from confidence output, not from frontmatter
• Stale memories identified and reported
• Health report generated (if requested)
• Any memory edits written to the memory body, not to frontmatter

References

• examples.md - Usage examples and workflows
• confidence-scoring.md - How confidence is calculated
• ADR-007 .project-toolkit/architecture/ADR-007-memory-first-architecture.md - Memory-first architecture
• ADR-038 .project-toolkit/architecture/ADR-038-reflexion-memory-schema.md - Reflexion memory schema