Baseline — ai-memory-mcp v0.6.3.1

Purpose. This file is the canonical, code-derived snapshot of what ai-memory-mcp v0.6.3.1 actually is — produced from main via 6 parallel read-only scans of the source tree. It is the reference doc that the long-running #512 drift audit measures published surfaces against.

Version naming. This product ships as release tag v0.6.3.1. The Cargo.toml version field encodes the same release as 0.6.3+patch.1 using SemVer build metadata — crates.io rejects 4-segment versions so the +patch.N suffix is the intentional encoding. Docs, HTML, and release notes always use the release-tag form v0.6.3.1. The Cargo encoding is internal.

Scope. Code is the only source of truth. This baseline does not contain any aspirational claims, marketing copy, or numbers without a traceable source.

Snapshot. main @ 9e85d36 (2026-05-01).

1. Workspace, crates, features, toolchain

1.1 Workspace identity

• Layout: single-crate repo (no [workspace] table). The root crate is monorepo-organized by convention.
• Root crate: ai-memory — Cargo.toml:1.
• Companion crate: ai-memory-fuzz — fuzz/Cargo.toml:2 (publish=false, fuzz harness only).

1.2 Versions (canonical = release tag)

| Crate | Cargo.toml | Release-tag form | | — | — | — | | ai-memory | 0.6.3+patch.1 (Cargo.toml:3) | v0.6.3.1 | | ai-memory-fuzz | 0.0.0 (fuzz/Cargo.toml:3) | n/a (unpublished) |

1.3 Targets

• Library: ai_memory (root, implicit from package.name).
• Binary: ai-memory (src/main.rs).
• Bench target: recall (Cargo.toml:198–200, harness = false).
• Fuzz binaries: fuzz_validate (fuzz/Cargo.toml:13), fuzz_namespace (fuzz/Cargo.toml:20).

1.4 Cargo features (root crate)

Source: Cargo.toml:134–153.

1.5 Optional dependencies

| Dep | Version | Gate | Cargo.toml | | — | — | — | — | | async-trait | 0.1 | sal | :116 | | bitflags | 2 | sal | :117 | | thiserror | 2 | sal | :118 | | sqlx | 0.8 | sal-postgres | :124 | | pgvector | 0.4 | sal-postgres | :125 |

1.6 Toolchain

• Edition (both crates): 2024.
• MSRV (root): 1.88 (Cargo.toml:5 rust-version = "1.88").
• MSRV (fuzz): undeclared.
• rust-toolchain.toml: not present.

2. MCP tool surface (43 tools at v0.6.3.1 baseline)

Forward note (v0.7.0): the 43-tool baseline below is the historical v0.6.3.1 snapshot. At v0.7.0 the surface advertises 74 entries at --profile full (73 callable memory tools + the always-on memory_capabilities bootstrap). Default --profile core (introduced at v0.6.4) ships 7 tools at v0.7.0 plus the always-on bootstrap. See CHANGELOG.md [Unreleased]/[0.7.0] for the full delta, and MIGRATION_v0.7.md for the upgrade path. The 43 below remain accurate as the v0.6.3.1 baseline this doc was captured at.

All registered in src/mcp.rs::tool_definitions() (lines 147–715) at v0.6.3.1; v0.7.0 moves this to src/mcp/registry.rs::tool_definitions() with per-Family filtering via tool_definitions_for_profile(). Dispatched in handle_request() (lines 3473–3572 at v0.6.3.1; now in src/mcp/mod.rs). Listed via tools/list handler at :3404–3445 at v0.6.3.1. The v0.6.3.1 schema-version stamp was toolsVersion = "2026-04-26" (src/mcp.rs:153); v0.7.0 capabilities envelope schema_version is "3".

No tool is #[cfg]-gated at compile time; tier-gated tools degrade at runtime to a 501 / "LLM not available" style response.

Aliases / deprecated names: none. The codebase uses exact string matching on tool name; no compatibility shims.

MCP prompts (non-tool resources):

• recall-first (src/mcp.rs:724) — system prompt for proactive memory recall, TOON format, tier strategy. Optional namespace argument.
• memory-workflow (src/mcp.rs:735) — quick reference card for tool usage patterns.

3. CLI surface

Single binary: ai-memory (src/main.rs → daemon_runtime::run() at src/daemon_runtime.rs:429). Uses clap derive.

3.1 Global flags (apply to every subcommand)

Source: src/daemon_runtime.rs:96–125.

3.2 Subcommand inventory (40 top-level)

Source: src/daemon_runtime.rs:127–259 and src/cli/*.rs.

serve, mcp, store, update, recall, search, get, list, delete, promote, forget, link, consolidate, gc, stats, namespaces, export, import, resolve, shell, sync, sync-daemon, auto-consolidate, completions, man, mine, archive, agents, pending, backup, restore, curator, bench, migrate (gated --features sal), doctor, boot, install, wrap, logs, audit.

Hidden subcommands: none (#[command(hide = true)] not used).

Nested subcommand groups:

• archive: list, restore, purge, stats.
• agents: list (default), register.
• pending: list, approve, reject.
• install: targets claude-code, openclaw, cursor, cline, continue, windsurf.
• logs: tail, cat, archive, purge.
• audit: verify, tail, path.

Per-subcommand flag detail is exhaustive and lives in src/cli/*.rs. Cross-check against docs/CLI_REFERENCE.md is the responsibility of the drift scanner.

3.3 Environment variables read

| Env var | Read site | Purpose | | — | — | — | | AI_MEMORY_DB | daemon_runtime.rs:105 | DB path (overrides --db) | | AI_MEMORY_AGENT_ID | daemon_runtime.rs:113, identity.rs:127 | Agent identifier | | AI_MEMORY_DB_PASSPHRASE | db.rs:206, daemon_runtime.rs:438 | SQLCipher passphrase (set internally from --db-passphrase-file) | | AI_MEMORY_ANONYMIZE | daemon_runtime.rs:881, config.rs:1309, identity.rs:54 | Anonymize agent_id | | AI_MEMORY_NO_CONFIG | config.rs:1220 | Bypass config.toml load | | AI_MEMORY_BOOT_ENABLED | config.rs:1139 | Boot feature gate | | AI_MEMORY_AUTONOMOUS_HOOKS | config.rs:1294 | Enable autonomous hooks | | AI_MEMORY_AUDIT_DIR | cli/audit.rs | Audit log directory | | AI_MEMORY_LOG_DIR | cli/logs.rs | Operational log directory | | HOME, XDG_STATE_HOME, LOCALAPPDATA, USERPROFILE | log_paths.rs, config.rs, embeddings.rs, audit.rs | Platform-dependent path resolution | | INVOCATION_ID | log_paths.rs:193 | systemd journal context detection | | PATH | cli/install.rs:365 | Installer binary lookup | | AI_MEMORY_TEST_POSTGRES_URL | store/postgres.rs:797 | Test-only Postgres URL (gated feature) |

4. Configuration (config.toml)

Loader: AppConfig::load() (src/config.rs:1232). Default location: ~/.config/ai-memory/config.toml.

4.1 Top-level keys ([AppConfig], src/config.rs:928–984)

| Key | Type | Default | Purpose | | — | — | — | — | | tier | Option<String> | "semantic" (resolved by effective_tier()) | Feature tier: keyword | semantic | smart | autonomous | | db | Option<String> | "ai-memory.db" | SQLite path | | ollama_url | Option<String> | "http://localhost:11434" | Ollama base URL | | embed_url | Option<String> | falls back to ollama_url | Separate embedding-model URL | | embedding_model | Option<String> | tier-derived | mini_lm_l6_v2 | nomic_embed_v15 | | llm_model | Option<String> | tier-derived | Ollama tag (e.g. gemma4:e2b) | | cross_encoder | Option<bool> | tier-derived | Enable neural reranker | | default_namespace | Option<String> | "global" | Default namespace for stores | | max_memory_mb | Option<usize> | none | Memory budget for auto-tier selection | | archive_on_gc | Option<bool> | true | Archive expired memories before GC delete | | api_key | Option<String> | none | HTTP X-API-Key | | archive_max_days | Option<i64> | none (disabled) | Archive retention horizon | | ttl | Option<TtlConfig> | compiled defaults | Per-tier TTL overrides | | scoring | Option<RecallScoringConfig> | compiled defaults | Recall time-decay half-life | | autonomous_hooks | Option<bool> | false | Fire LLM hooks on store | | logging | Option<LoggingConfig> | disabled | Operational logging | | audit | Option<AuditConfig> | disabled | Security audit trail | | boot | Option<BootConfig> | enabled, no redact | Boot privacy controls | | identity | Option<IdentityConfig> | none | Identity resolution |

4.2 Sub-tables

[ttl] (src/config.rs:727–738):

• short_ttl_secs — default 21600 (6h).
• mid_ttl_secs — default 604800 (7d).
• long_ttl_secs — default None (never expires).
• short_extend_secs — default 3600 (1h, on-access TTL bump).
• mid_extend_secs — default 86400 (1d, on-access TTL bump).

[logging] (src/config.rs:992–1015):

• enabled (false), path (~/.local/state/ai-memory/logs/), max_size_mb (100), max_files (30), retention_days (90), structured (false), level (info), rotation (daily), filename_prefix (ai-memory.log).

[audit] (src/config.rs:1024–1061):

• enabled (false), path (~/.local/state/ai-memory/audit/), schema_version (validated against binary), redact_content (v1 only supports true), hash_chain (true), attestation_cadence_minutes (60), append_only (true), retention_days (90), compliance (sub-table for SOC2/HIPAA/GDPR/FedRAMP presets).

[boot] (src/config.rs:1118–1130):

• enabled (true; env AI_MEMORY_BOOT_ENABLED), redact_titles (false).

[identity] (src/config.rs:1202–1208):

• anonymize_default (false).

5. Storage: schema, migrations, tiers

5.1 SQLite schema

• Current schema version: 19 (src/db.rs:178, CURRENT_SCHEMA_VERSION).
• Migration files: migrations/sqlite/
  • 0010_v063_hierarchy_kg.sql → v15 (KG temporal validity + entity_aliases).
  • 0011_v0631_data_integrity.sql → v18 (embedding_dim guard, archive lossless).
  • 0012_governance_inherit.sql → v17 (governance.inherit backfill).
  • 0013_webhook_event_types.sql → v19 (subscriptions event_types column + index).

5.2 Tables (canonical shape, post-migration v19)

• memories (src/db.rs:111–127) — primary store. Columns: id, tier, namespace, title, content, tags (JSON), priority, confidence, source, access_count, created_at, updated_at, last_accessed_at, expires_at, metadata (JSON), embedding (BLOB), embedding_dim, generated columns scope_idx and agent_id_idx. UNIQUE(title, namespace).
• memory_links (src/db.rs:135–141) — KG edges. PK (source_id, target_id, relation). Columns: relation, created_at, valid_from, valid_until, observed_by, signature. Temporal-validity indexes on (source_id, valid_from, valid_until) and (target_id, valid_from, valid_until) and (relation, valid_from).
• memories_fts (src/db.rs:143–166) — FTS5 virtual table mirroring title/content/tags. INSERT/DELETE/UPDATE triggers (ai, ad, au).
• archived_memories (src/db.rs:306–327) — same shape as memories plus archived_at, archive_reason (default 'ttl_expired'), original_tier (default 'long' for pre-v18), original_expires_at.
• entity_aliases (migrations/sqlite/0010:26–34) — PK (entity_id, alias); created_at non-null; index on alias.
• namespace_meta (src/db.rs:331–336) — PK namespace; standard_id (FK → memories), updated_at, parent_namespace.
• pending_actions (src/db.rs:371–385) — governance queue. Status pending|approved|rejected; approvals JSON array.
• sync_state (src/db.rs:446–454) — federation high-watermarks. PK (agent_id, peer_id).
• subscriptions (src/db.rs:478–496) — webhook subs. event_types (JSON) + index added in v19.
• schema_version (src/db.rs:168–170) — single integer column.

5.3 Postgres adapter

• Current schema version (Postgres): 15 (src/store/postgres.rs:65). Lags SQLite (19).

5.4 Memory tiers (Tier enum)

Source: src/models.rs:10–51.

5.5 Feature tiers (orthogonal capability hierarchy)

Source: FeatureTier enum, src/config.rs:79–161.

Memory-tier (Short/Mid/Long) is orthogonal to feature-tier (keyword/semantic/smart/autonomous).

6. Hardcoded product-facing constants (63 total)

Source-of-truth numbers that must match docs or be flagged. Categories below; each row cites file:line.

6.1 Capacity / size limits

| Constant | Value | Location | | — | — | — | | MAX_CONTENT_SIZE | 65,536 bytes | src/models.rs:859 | | MAX_TITLE_LEN | 512 chars | src/validate.rs:11 | | MAX_NAMESPACE_LEN | 512 chars | src/validate.rs:15 | | MAX_NAMESPACE_DEPTH | 8 levels | src/models.rs:863 | | MAX_METADATA_SIZE | 65,536 bytes | src/validate.rs:22 | | MAX_METADATA_DEPTH | 32 levels | src/validate.rs:23 | | MAX_TAG_LEN | 128 chars | src/validate.rs:17 | | MAX_TAGS_COUNT | 50 items | src/validate.rs:18 | | MAX_SOURCE_LEN | 64 chars | src/validate.rs:16 | | MAX_RELATION_LEN | 64 chars | src/validate.rs:19 | | MAX_ID_LEN | 128 chars | src/validate.rs:20 | | MAX_AGENT_ID_LEN | 128 chars | src/validate.rs:21 | | MAX_BULK_SIZE | 1,000 items | src/handlers.rs:69 | | HNSW MAX_ENTRIES | 100,000 vectors | src/hnsw.rs:20 | | MAX_VERSION_SUFFIX (on_conflict=version) | 1,024 versions | src/db.rs:1755 | | TAXONOMY_MAX_LIMIT | 10,000 nodes | src/db.rs:2085 | | MAX_ROWS (migration ceiling) | 1,000,000 memories | src/migrate.rs:137 |

6.2 Time / TTL / retention

| Constant | Value | Location | Override | | — | — | — | — | | Short tier default TTL | 21,600 s (6h) | src/models.rs:46 | [ttl].short_ttl_secs | | Mid tier default TTL | 604,800 s (7d) | src/models.rs:47 | [ttl].mid_ttl_secs | | Long tier default TTL | none (∞) | src/models.rs:48 | [ttl].long_ttl_secs | | SHORT_TTL_EXTEND_SECS | 3,600 s | src/models.rs:946 | [ttl].short_extend_secs | | MID_TTL_EXTEND_SECS | 86,400 s | src/models.rs:947 | [ttl].mid_extend_secs | | MAX_TTL_SECS (clamp ceiling) | 315,360,000 s (~10y) | src/config.rs:765 | none | | GC_INTERVAL_SECS (daemon) | 1,800 s (30 min) | src/daemon_runtime.rs:83 | none | | WAL_CHECKPOINT_INTERVAL_SECS | 600 s (10 min) | src/daemon_runtime.rs:86 | none | | Curator default interval | 3,600 s (1h) | src/curator.rs:36 | [curator].interval_secs |

6.3 Batch / chunk sizes

| Constant | Value | Location | | — | — | — | | BULK_FANOUT_CONCURRENCY | 8 | src/handlers.rs:78 | | HNSW REBUILD_THRESHOLD | 200 overflow entries | src/hnsw.rs:17 | | memory_recall default limit | 50 (clamp [1,1000]) | src/handlers.rs:3089 | | memory_list default limit | 50 (clamp [1,1000]) | src/handlers.rs:3089 | | memory_inbox default limit | 10 (clamp [1,50]) | src/handlers.rs:1568 | | get_taxonomy default limit | 1,000 (clamp [1,10000]) | src/handlers.rs:1998 | | KG_TIMELINE_DEFAULT_LIMIT | 200 | src/db.rs:2623 | | KG_TIMELINE_MAX_LIMIT | 1,000 | src/db.rs:2627 | | KG_QUERY_DEFAULT_LIMIT | 200 | src/db.rs:2781 | | KG_QUERY_MAX_LIMIT | 1,000 | src/db.rs:2786 |

6.4 Timeouts / intervals

| Constant | Value | Location | | — | — | — | | GENERATE_TIMEOUT (LLM call) | 30 s | src/llm.rs:10 | | PULL_TIMEOUT (Ollama model pull) | 120 s | src/llm.rs:11 | | FANOUT_RETRY_BACKOFF | 250 ms | src/federation.rs:382 | | Postgres DEFAULT_ACQUIRE_TIMEOUT | 30 s | src/store/postgres.rs:71 | | Federation handshake connect_timeout | 2 s | src/federation.rs:145 |

6.5 Tier boundaries / similarity thresholds

| Constant | Value | Location | | — | — | — | | PROMOTION_THRESHOLD (access count) | 5 | src/models.rs:944 | | DUPLICATE_THRESHOLD_MIN (cosine) | 0.5 | src/db.rs:2287 | | DUPLICATE_THRESHOLD_DEFAULT (cosine) | 0.85 | src/db.rs:2293 | | CONSOLIDATE_JACCARD_THRESHOLD | 0.55 | src/autonomy.rs:45 | | CONSOLIDATE_MAX_CLUSTER_SIZE | 8 | src/autonomy.rs:49 |

6.6 Concurrency

| Constant | Value | Location | | — | — | — | | Postgres DEFAULT_MAX_CONNECTIONS | 16 | src/store/postgres.rs:70 |

6.7 Knowledge graph

| Constant | Value | Location | | — | — | — | | KG_QUERY_MAX_SUPPORTED_DEPTH | 5 hops | src/db.rs:2792 | | MAX_EXPLICIT_DEPTH (legacy contradicts-search cycle limit) | 8 hops | src/db.rs:4384 | | MAX_DEPTH (histogram bucket ceiling) | 16 hops | src/db.rs:5141 |

6.8 Embedding / reranking

| Constant | Value | Location | | — | — | — | | MINILM_DIM | 384 | src/embeddings.rs:16 | | MAX_SEQ_LEN (MiniLM) | 256 tokens | src/embeddings.rs:17 | | EMBEDDING_DIM (public re-export) | 384 | src/embeddings.rs:285 | | NOMIC_DIM | 768 | src/embeddings.rs:25 | | CROSS_ENCODER_MAX_SEQ | 512 tokens | src/reranker.rs:33 | | CROSS_ENCODER_WEIGHT | 0.4 | src/reranker.rs:30 | | ORIGINAL_WEIGHT (FTS+HNSW blend) | 0.6 | src/reranker.rs:28 |

6.9 Schema-version range

| Constant | Value | Location | | — | — | — | | MIN_SUPPORTED_SCHEMA | 16 | src/cli/boot.rs:54 | | MAX_SUPPORTED_SCHEMA | 19 | src/cli/boot.rs:61 | | SQLite CURRENT_SCHEMA_VERSION | 19 | src/db.rs:178 | | Postgres CURRENT_SCHEMA_VERSION | 15 | src/store/postgres.rs:65 |

6.10 Misc product-facing

| Constant | Value | Location | | — | — | — | | DEFAULT_PORT (HTTP daemon) | 9077 | src/daemon_runtime.rs:82 | | Boot hook DEFAULT_BUDGET_TOKENS | 4,096 | src/cli/boot.rs:76 | | TOKENS_PER_CHAR (boot estimator, advisory) | 0.25 | src/cli/boot.rs:82 | | Wrap DEFAULT_BUDGET_TOKENS | 4,096 | src/cli/wrap.rs:61 | | Wrap DEFAULT_LIMIT | 10 | src/cli/wrap.rs:65 | | Curator DEFAULT_MAX_OPS_PER_CYCLE | 100 | src/curator.rs:39 | | Curator MIN_CONTENT_LEN | 50 chars | src/curator.rs:43 |

User-overridable subset: 16 of 63 constants are reachable via config.toml, CLI flags, or query parameters. The rest are compile-time hard limits.

7. Performance: benches and published claims

7.1 Bench inventory

Inline (ai-memory bench subcommand) — src/bench.rs:

• StoreNoEmbedding — memory_store without embedder (pure SQLite write).
• SearchFts — memory_search via FTS5 (200-memory corpus).
• RecallHot — memory_recall hot path, depth=1 (200-memory corpus).
• KgQueryDepth1 — memory_kg_query depth=1 (fan-out: 50 sources × 4 links).
• KgQueryDepth3 — memory_kg_query depth=3 (chain: 50 × 5 hops).
• KgQueryDepth5 — memory_kg_query depth=5 tail-case.
• KgTimeline — memory_kg_timeline ordered fact timeline.

Criterion (benches/recall.rs):

• bench_recall: short_query, medium_query, long_query (against 1000 seeded memories).
• bench_search: simple_search, filtered_search.
• bench_insert: store_memory.

Cargo [[bench]]: one entry, name recall, harness = false (Cargo.toml:198–200).

Committed bench-result outputs: none. target/criterion is gitignored. benchmarks/longmemeval/ exists (Python harness) but holds no committed outputs.

7.2 Published claims vs. backing

Total numeric claims across PERFORMANCE.md and docs/performance.html: 29. Backed by an inline bench: 7. UNBACKED (no bench produces this number): 22.

PERFORMANCE.md:76 notes embedder-bound paths as 🚧 Stream E follow-up. PERFORMANCE.md:124–129 notes curator/federation/CLI as “not yet wired in.” 22 of 29 claims have no temporal commitment for adding a matching bench.

8. CI / release-engineering surface

Workflows under .github/workflows/:

Implication for the drift workflow: docs-only PRs to chore/product-drift-audit and merges to main do not initiate a release. The release path is git tag v* && git push --tags exclusively.

9. Generation provenance

This baseline was generated on 2026-05-01 from the working tree at main (9e85d36) via 6 parallel read-only Explore agents covering: workspace/features, MCP tool surface, CLI surface, config + schema + tiers, hardcoded constants, benches + perf claims. Every fact above cites file:line against the source tree. No file was modified during generation.

The drift scanner (issue #512, weekly) measures published surfaces against this baseline. All resolutions are documentation-only — code is the source of truth and is never edited by the drift workflow.