brain-memory 0.1.0-beta.33 → 0.1.0-beta.35

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -6,6 +6,85 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/), and this
6
6
 
7
7
  ## [Unreleased]
8
8
 
9
+ ## [0.1.0-beta.35] - 2026-08-01
10
+
11
+ ### Added
12
+
13
+ - **`/brain:import` — cold-start a brain from transcripts your agents already
14
+ wrote.** A new brain is empty, and an empty brain is worth nothing until
15
+ something fills it; meanwhile every coding agent has been keeping months of
16
+ local history. `brain import` harvests it: session titles, user prompts,
17
+ projects, branches and edited files, with harness wrappers
18
+ (`<system-reminder>`, slash-command tags, command output), subagent sidechain
19
+ traffic, and bare acknowledgements filtered out. The digest is bounded on
20
+ three axes (sessions, prompts per session, total characters) and
21
+ round-robins across projects, so one busy repo cannot consume the whole
22
+ budget — on a real 261-session history that lifted project coverage from 5
23
+ to 13 within the same ~15k tokens.
24
+
25
+ The CLI harvests; the **agent distills**. Extracting meaning is a semantic
26
+ judgement, and making it in the CLI would mean shipping an embedding model
27
+ or calling an LLM — the two things Brain exists to avoid. So the harvester
28
+ reports facts and never infers, and distilled memories go through the
29
+ ordinary `brain memorize` path as `agent-inferred`: an import inherits the
30
+ existing provenance ceilings for free and can never pin, entrench, or
31
+ outrank something the user said directly.
32
+
33
+ Import is incremental — a cursor at `~/.brain/.import/state.json` records
34
+ which sessions have been read so re-runs offer only new ones (`--all`
35
+ overrides). `--mark` validates ids against the source's real sessions and
36
+ exits non-zero on any that match nothing: silently accepting a truncated or
37
+ mistyped id would report a session as retired while the cursor never matches
38
+ it, re-offering that session forever. Scope with `--project`,
39
+ `--since 30d|6m|2026-01-01`, `--limit`.
40
+ `brain import --sources` lists detected history stores. `brain import` only
41
+ ever reads and prints; no memory is written without the agent.
42
+
43
+ Claude Code is the first source adapter; the registry in `src/harvest.js`
44
+ takes additional agents without changes downstream.
45
+
46
+ ## [0.1.0-beta.34] - 2026-07-26
47
+
48
+ ### Added
49
+
50
+ - **Memory provenance — every write is now labelled by origin.** Each memory
51
+ carries an `origin` (`user`, `agent-inferred`, `tool-output`, `external`)
52
+ recording *where the content came from*, which is a different question from
53
+ how much the agent believes it. Origin decides what a memory may claim:
54
+ non-user origins are capped below the 0.7 prune-exempt salience threshold,
55
+ capped on confidence so they stay flagged as uncertain at recall, may not
56
+ raise their own base strength, and decay faster — so a fact absorbed from
57
+ untrusted content fades and loses to a genuine one over time. Defaults to
58
+ `agent-inferred` when absent (the safe direction). Downgrades are never
59
+ silent: `brain memorize` reports every value it lowered under
60
+ `provenance_clamps`. Mitigates OWASP ASI06 (memory poisoning); see
61
+ MemGhost, [arXiv:2607.05189](https://arxiv.org/abs/2607.05189).
62
+ - **Append-only provenance log** at `~/.brain/audit.log` — one JSON object per
63
+ write, recorded before the caller sees success, so a fact that later proves
64
+ to be planted can be traced to the write that introduced it even if the
65
+ memory was since edited, consolidated, or deleted.
66
+
67
+ ### Changed
68
+
69
+ - **Breaking (CLI):** born-pinning through `brain memorize` now requires
70
+ `origin: "user"`. Entrenchment — `pinned` (loaded into every session) and
71
+ `stable` (exempt from decay) — is a capability rather than a magnitude, so a
72
+ request from any other origin is refused with a non-zero exit rather than
73
+ quietly capped. Pin deliberately afterwards with `brain pin <id>` instead.
74
+ Only affects callers that drive the CLI directly; `/brain:memorize` already
75
+ proposed rather than assumed pinning.
76
+
77
+ ### Fixed
78
+
79
+ - **Cache-honest token accounting in the benchmark harness.** `normalizeUsage`
80
+ now folds `cache_read`/`cache_creation` input tokens back into the input
81
+ count — the cached prefix was being undercounted, which deflated
82
+ tokens-per-successful-task on the DeepSeek endpoint. Cached tokens are
83
+ accumulated, logged, and persisted separately so the two can be compared.
84
+ Also: model-aware judge dispatch, `runs_per_scenario` raised to 10, tighter
85
+ needle-fact rubrics on scenarios A/B/C, and three amendment rows recording
86
+ all of it in `PREREGISTRATION.md`.
87
+
9
88
  ## [0.1.0-beta.33] - 2026-07-04
10
89
 
11
90
  _Covers everything since v0.1.0-beta.31, including the changes first shipped in the changelog-less 0.1.0-beta.32._
package/README.md CHANGED
@@ -242,10 +242,27 @@ The everyday loop is **ambient** — `remember` and `memorize` run automatically
242
242
  | `/brain:status` | Dashboard with brain health metrics and recommendations |
243
243
  | `/brain:pin [id\|query]` | Pin a memory to the always-present tier — loads every session, never decays. Toggles: also unpins (`--off`) |
244
244
  | `/brain:forget [target]` | Decay, archive, or remove memories. `--deep` performs forensic erasure — traces and removes every reference |
245
+ | `/brain:import [--project P] [--since 30d]` | Cold-start a new brain from transcripts your agents already wrote. Incremental — safe to re-run |
245
246
  | `/brain:sync [subcommand]` | Sync via Brain Cloud, Git remote, or export/import (auto-initializes the brain on first run) |
246
247
  | `/brain:skills [list\|show\|add\|use\|remove\|export]` | Manage procedural skills — reusable how-to workflows with progressive disclosure |
247
248
  | `/brain:sleep [scope]` | Full maintenance cycle — 9 neuroscience-inspired phases (replay, consolidation, review reinforcement, pruning, dreaming, …). Usually runs automatically/in the background |
248
249
 
250
+ ## Cold Start
251
+
252
+ A new brain is empty, and an empty brain is worth nothing until something fills it. But your agents have been keeping local transcripts for months — decisions, preferences, corrections, conventions. `/brain:import` reads them.
253
+
254
+ ```bash
255
+ brain import --sources # what history exists on this machine
256
+ brain import --project my-app # digest that project's past sessions
257
+ brain import --since 30d # or scope by time
258
+ ```
259
+
260
+ The split matters. The CLI **harvests** — deterministic, budget-bounded, no model call: it extracts session titles, user prompts, projects, branches, and edited files, filters out harness noise and subagent traffic, and interleaves across projects so one busy repo can't eat the budget. The agent then **distills** that digest into memories through the ordinary `brain memorize` path.
261
+
262
+ That split is deliberate. Extracting meaning is a semantic judgement, and doing it in the CLI would mean shipping an embedding model or calling an LLM — the two things Brain exists to avoid. So the harvester reports facts and never infers, and every imported memory lands as `agent-inferred` provenance: it inherits the existing ceilings for free, and **an import can never pin, entrench, or outrank something you said directly.**
263
+
264
+ Import is incremental. A cursor in `~/.brain/.import/state.json` records which sessions have been read, so re-running only offers new ones (`--all` overrides). Nothing is written without you — `brain import` only ever reads and prints.
265
+
249
266
  ## Session Lifecycle
250
267
 
251
268
  Brain Memory works automatically in the background — no commands needed for basic context awareness.
package/bin/brain.js CHANGED
@@ -33,6 +33,7 @@ const DELEGATED = {
33
33
  pin: 'pin.js',
34
34
  unpin: 'unpin.js',
35
35
  skill: 'skill.js',
36
+ import: 'import.js',
36
37
  };
37
38
 
38
39
  // install.js finds its positional subcommand from process.argv itself, so it is
@@ -59,6 +60,13 @@ Memory
59
60
  session-start [--project P] [--task T] [--top N]
60
61
  Budget-bounded startup payload (agent-invoked)
61
62
 
63
+ Cold start
64
+ import [--source S] [--project P] [--since 30d] [--limit N] [--all]
65
+ Digest past agent transcripts into import candidates
66
+ import --sources List detected agent history stores
67
+ import --mark <id> [<id>...]
68
+ Record sessions as imported (incremental cursor)
69
+
62
70
  Sync
63
71
  cloud <login|logout|push|pull|status>
64
72
 
package/bin/import.js ADDED
@@ -0,0 +1,132 @@
1
+ #!/usr/bin/env node
2
+
3
+ /**
4
+ * brain import — cold-start a brain from transcripts your agents already wrote.
5
+ *
6
+ * Emits a deterministic, budget-bounded digest of past sessions. The agent reads
7
+ * that digest and writes memories through the ordinary `brain memorize` path, so
8
+ * imported memories go through the same provenance gate, scoring, and index as
9
+ * anything else. Nothing here writes a memory.
10
+ *
11
+ * Usage:
12
+ * brain import [--source claude-code] [--project P] [--since ISO] [--limit N] [--all]
13
+ * brain import --sources List detected history stores
14
+ * brain import --mark <id> [<id>...] Record sessions as imported
15
+ *
16
+ * The `--mark` step is what makes import incremental: the agent calls it after a
17
+ * successful memorize so the next run offers only new sessions.
18
+ */
19
+
20
+ const fs = require('fs');
21
+ const path = require('path');
22
+
23
+ const { getBrainDir } = require('../src/index-manager');
24
+ const { harvest, availableSources, markImported } = require('../src/harvest');
25
+
26
+ function parseArgs(argv) {
27
+ const args = {
28
+ source: 'claude-code',
29
+ project: null,
30
+ since: null,
31
+ limit: null,
32
+ all: false,
33
+ sources: false,
34
+ mark: [],
35
+ };
36
+ for (let i = 0; i < argv.length; i++) {
37
+ switch (argv[i]) {
38
+ case '--source': args.source = argv[++i]; break;
39
+ case '--project': args.project = argv[++i]; break;
40
+ case '--since': args.since = argv[++i]; break;
41
+ case '--limit': args.limit = parseInt(argv[++i], 10) || null; break;
42
+ case '--all': args.all = true; break;
43
+ case '--sources': args.sources = true; break;
44
+ case '--mark':
45
+ // Consume every following non-flag token as a session id.
46
+ while (argv[i + 1] && !argv[i + 1].startsWith('--')) args.mark.push(argv[++i]);
47
+ break;
48
+ default: break;
49
+ }
50
+ }
51
+ return args;
52
+ }
53
+
54
+ /**
55
+ * Normalize `--since` into an ISO timestamp.
56
+ *
57
+ * Accepts a relative shorthand (`30d`, `6m`, `1y`) as well as any date string
58
+ * Date can parse, because "the last month of work" is the natural way to scope
59
+ * an import and computing that date by hand is friction.
60
+ */
61
+ function normalizeSince(value) {
62
+ if (!value) return null;
63
+ const relative = /^(\d+)([dmy])$/i.exec(value.trim());
64
+ if (relative) {
65
+ const amount = parseInt(relative[1], 10);
66
+ const date = new Date();
67
+ if (relative[2].toLowerCase() === 'd') date.setDate(date.getDate() - amount);
68
+ if (relative[2].toLowerCase() === 'm') date.setMonth(date.getMonth() - amount);
69
+ if (relative[2].toLowerCase() === 'y') date.setFullYear(date.getFullYear() - amount);
70
+ return date.toISOString();
71
+ }
72
+ const parsed = new Date(value);
73
+ if (Number.isNaN(parsed.getTime())) return null;
74
+ return parsed.toISOString();
75
+ }
76
+
77
+ function main(argv = process.argv.slice(2)) {
78
+ const args = parseArgs(argv);
79
+
80
+ if (args.sources) {
81
+ console.log(JSON.stringify({ sources: availableSources() }, null, 2));
82
+ return;
83
+ }
84
+
85
+ // A brain must exist before we offer to fill it — otherwise the agent gets a
86
+ // digest it has nowhere to write.
87
+ const brainDir = getBrainDir();
88
+ if (!fs.existsSync(path.join(brainDir, 'index.json'))) {
89
+ console.error(JSON.stringify({ error: 'Brain not initialized. Run `brain install` first.' }));
90
+ process.exit(1);
91
+ }
92
+
93
+ if (args.mark.length > 0) {
94
+ const result = markImported(args.source, args.mark);
95
+ console.log(JSON.stringify({
96
+ marked: args.mark.length - result.unknown.length,
97
+ ...result,
98
+ source: args.source,
99
+ ...(result.unknown.length
100
+ ? { warning: `${result.unknown.length} id(s) matched no ${args.source} session and were not recorded — check for a truncated or mistyped session id.` }
101
+ : {}),
102
+ }, null, 2));
103
+ // A partial mark is a real failure: the caller believes those sessions are
104
+ // done, and nothing else will ever tell them otherwise.
105
+ if (result.unknown.length > 0) process.exit(1);
106
+ return;
107
+ }
108
+
109
+ if (args.since && !normalizeSince(args.since)) {
110
+ console.error(JSON.stringify({ error: `Could not parse --since "${args.since}" (try 30d, 6m, or 2026-01-01)` }));
111
+ process.exit(1);
112
+ }
113
+
114
+ const digest = harvest({
115
+ source: args.source,
116
+ project: args.project,
117
+ since: normalizeSince(args.since),
118
+ limit: args.limit,
119
+ all: args.all,
120
+ });
121
+
122
+ if (digest.error) {
123
+ console.error(JSON.stringify({ error: digest.error }));
124
+ process.exit(1);
125
+ }
126
+
127
+ console.log(JSON.stringify(digest, null, 2));
128
+ }
129
+
130
+ if (require.main === module) main();
131
+
132
+ module.exports = { main, parseArgs, normalizeSince };
package/bin/memorize.js CHANGED
@@ -71,6 +71,38 @@ const COGNITIVE_ADJUSTMENTS = {
71
71
  procedural: { strength_delta: -0.10, decay_multiplier: 1.003 },
72
72
  };
73
73
 
74
+ // --- Provenance (OWASP ASI06: memory poisoning) ---
75
+ //
76
+ // Where a memory came from decides how much it is allowed to entrench itself.
77
+ // The failure mode this defends against: content the agent *read* (an email, a
78
+ // web page, a tool result) persuades it to write a fact, and that fact then
79
+ // hardens — pinned into every session, exempt from decay, immune to pruning,
80
+ // and never flagged as uncertain. See MemGhost (arXiv:2607.05189).
81
+ //
82
+ // `origin` is asserted by the caller, so it does not defend against a fully
83
+ // hostile agent — but the dominant real case is an *honest* agent relaying
84
+ // poisoned content, and there it holds. The pin/stable restriction is the part
85
+ // that holds regardless: entrenchment is simply not reachable from this path
86
+ // without an explicit user origin.
87
+ const ORIGIN_POLICY = {
88
+ // The user asked for this directly, in-session.
89
+ user: { max_salience: 1.0, max_confidence: 1.0, allow_entrench: true, decay_multiplier: 1.0 },
90
+ // The agent inferred or summarized it from session context. Default.
91
+ 'agent-inferred': { max_salience: 0.6, max_confidence: 0.8, allow_entrench: false, decay_multiplier: 1.0 },
92
+ // Derived from tool output — file reads, command results, MCP responses.
93
+ 'tool-output': { max_salience: 0.5, max_confidence: 0.6, allow_entrench: false, decay_multiplier: 0.997 },
94
+ // Derived from untrusted external content — email, web pages, issue text.
95
+ external: { max_salience: 0.4, max_confidence: 0.4, allow_entrench: false, decay_multiplier: 0.99 },
96
+ };
97
+
98
+ // Absent origin means the agent didn't tell us — assume the weaker claim, never
99
+ // the stronger one. A memory that deserves `user` is one keystroke away.
100
+ const DEFAULT_ORIGIN = 'agent-inferred';
101
+
102
+ // Salience >= 0.7 is exempt from auto-pruning, so every non-user ceiling sits
103
+ // below it: an unattended poisoned memory must remain collectable.
104
+ const PRUNE_EXEMPT_SALIENCE = 0.7;
105
+
74
106
  // --- Args ---
75
107
 
76
108
  function parseArgs(argv) {
@@ -83,7 +115,61 @@ function parseArgs(argv) {
83
115
 
84
116
  // --- Helpers ---
85
117
 
86
- function computeStrengthAndDecay(type, cognitiveType, strengthAdjustment = 0) {
118
+ /**
119
+ * Resolve a memory's origin and clamp its trust-bearing fields to what that
120
+ * origin is permitted to claim.
121
+ *
122
+ * Returns `{ origin, policy, mem, clamps }` where `mem` is a sanitized copy and
123
+ * `clamps` lists every explicit over-ask the policy lowered — surfaced in the
124
+ * CLI output so a downgrade is never silent. Returns `{ error }` for requests
125
+ * that have no benign reading.
126
+ */
127
+ function applyOriginPolicy(mem) {
128
+ const origin = mem.origin || DEFAULT_ORIGIN;
129
+ const policy = ORIGIN_POLICY[origin];
130
+ if (!policy) {
131
+ return { error: `Unknown origin "${origin}" (expected one of: ${Object.keys(ORIGIN_POLICY).join(', ')})` };
132
+ }
133
+
134
+ // Entrenchment — loading into every session, or exemption from decay — is a
135
+ // capability, not a magnitude. Tool output and inbound email have no benign
136
+ // reason to request it, so this is refused loudly rather than quietly capped.
137
+ if (!policy.allow_entrench && (mem.pinned || mem.stable)) {
138
+ const asked = [mem.pinned && 'pinned', mem.stable && 'stable'].filter(Boolean).join(' + ');
139
+ return {
140
+ error: `Origin "${origin}" may not set ${asked}. Entrenching a memory requires origin ` +
141
+ `"user" — or pin it deliberately afterwards with \`brain pin <id>\`.`,
142
+ };
143
+ }
144
+
145
+ const clamps = [];
146
+ const clamp = (field, requested, fallback, max) => {
147
+ const value = requested ?? fallback;
148
+ if (value <= max) return value;
149
+ // Only report a downgrade the caller actually asked for. A lowered default
150
+ // is the policy working as designed, not a rejected claim.
151
+ if (requested != null) clamps.push({ field, requested, allowed: max, origin });
152
+ return max;
153
+ };
154
+
155
+ return {
156
+ origin,
157
+ policy,
158
+ clamps,
159
+ mem: {
160
+ ...mem,
161
+ salience: clamp('salience', mem.salience, 0.5, policy.max_salience),
162
+ confidence: clamp('confidence', mem.confidence, 0.7, policy.max_confidence),
163
+ // A non-user origin may lower a memory's base strength but never raise it.
164
+ strength_adjustment: clamp(
165
+ 'strength_adjustment', mem.strength_adjustment, 0,
166
+ policy.allow_entrench ? Infinity : 0
167
+ ),
168
+ },
169
+ };
170
+ }
171
+
172
+ function computeStrengthAndDecay(type, cognitiveType, strengthAdjustment = 0, originDecayMultiplier = 1.0) {
87
173
  const typeDefaults = TYPE_DEFAULTS[type] || TYPE_DEFAULTS.observation;
88
174
  const cogAdj = COGNITIVE_ADJUSTMENTS[cognitiveType] || COGNITIVE_ADJUSTMENTS.semantic;
89
175
 
@@ -95,14 +181,20 @@ function computeStrengthAndDecay(type, cognitiveType, strengthAdjustment = 0) {
95
181
  // over time (strength * decay_rate^days) — inverted decay. 0.9999/day is
96
182
  // "extremely slow" (a memory loses ~3% of its strength per year) without ever
97
183
  // inverting.
98
- const decay_rate = Math.min(0.9999, typeDefaults.decay_rate * cogAdj.decay_multiplier);
184
+ // The origin multiplier rides on top: memory sourced from untrusted content
185
+ // fades faster, so a planted fact loses to a genuine one over time even if
186
+ // nothing ever detects it as an attack.
187
+ const decay_rate = Math.min(
188
+ 0.9999,
189
+ typeDefaults.decay_rate * cogAdj.decay_multiplier * originDecayMultiplier
190
+ );
99
191
 
100
192
  return { strength: Math.round(strength * 100) / 100, decay_rate };
101
193
  }
102
194
 
103
- function buildMemoryFileContent(mem, id, now) {
195
+ function buildMemoryFileContent(mem, id, now, origin, originDecayMultiplier) {
104
196
  const { strength, decay_rate } = computeStrengthAndDecay(
105
- mem.type, mem.cognitive_type, mem.strength_adjustment
197
+ mem.type, mem.cognitive_type, mem.strength_adjustment, originDecayMultiplier
106
198
  );
107
199
 
108
200
  const fmLines = [
@@ -130,6 +222,7 @@ function buildMemoryFileContent(mem, id, now) {
130
222
  fmLines.push(
131
223
  `tags: [${(mem.tags || []).map(t => `"${t}"`).join(', ')}]`,
132
224
  `related: [${(mem.related || []).map(r => `"${r}"`).join(', ')}]`,
225
+ `origin: "${origin}"`,
133
226
  `source: "${mem.source || ''}"`,
134
227
  `encoding_context:`,
135
228
  ` project: "${(mem.encoding_context && mem.encoding_context.project) || ''}"`,
@@ -144,7 +237,7 @@ function buildMemoryFileContent(mem, id, now) {
144
237
  return { fileContent: frontmatter + mem.content + '\n', strength, decay_rate };
145
238
  }
146
239
 
147
- function buildIndexEntry(mem, id, strength, decayRate, now) {
240
+ function buildIndexEntry(mem, id, strength, decayRate, now, origin) {
148
241
  const entry = {
149
242
  title: mem.title,
150
243
  path: mem.path,
@@ -155,6 +248,9 @@ function buildIndexEntry(mem, id, strength, decayRate, now) {
155
248
  access_count: 0,
156
249
  strength,
157
250
  decay_rate: decayRate,
251
+ // Provenance travels with the index entry so recall can weigh and flag it
252
+ // without having to open every memory file.
253
+ origin,
158
254
  salience: mem.salience ?? 0.5,
159
255
  confidence: mem.confidence ?? 0.7,
160
256
  tags: mem.tags || [],
@@ -173,6 +269,22 @@ function buildIndexEntry(mem, id, strength, decayRate, now) {
173
269
  return entry;
174
270
  }
175
271
 
272
+ /**
273
+ * Append-only provenance log, one JSON object per line.
274
+ *
275
+ * Written for every memory that reaches disk, including ones the policy
276
+ * downgraded, so a fact that later turns out to be planted can be traced to the
277
+ * write that introduced it — even if the memory file itself was since edited,
278
+ * consolidated by a sleep cycle, or deleted.
279
+ */
280
+ function appendAuditLog(brainDir, record) {
281
+ fs.appendFileSync(
282
+ path.join(brainDir, 'audit.log'),
283
+ JSON.stringify(record) + '\n',
284
+ { mode: 0o600 }
285
+ );
286
+ }
287
+
176
288
  function updateMetaFiles(brainDir, memPath) {
177
289
  const parts = memPath.split('/');
178
290
  // Walk up the directory chain, updating _meta.json at each level
@@ -327,27 +439,42 @@ async function main() {
327
439
  const results = [];
328
440
  const newIds = [];
329
441
  const pinnedToAdd = [];
442
+ const clampsReported = [];
443
+ const auditErrors = [];
330
444
 
331
- for (const mem of input.memories) {
445
+ for (const rawMem of input.memories) {
332
446
  // Validate required fields
333
- if (!mem.title || !mem.type || !mem.path || !mem.content) {
447
+ if (!rawMem.title || !rawMem.type || !rawMem.path || !rawMem.content) {
334
448
  console.error(JSON.stringify({
335
- error: `Memory missing required fields (title, type, path, content): ${JSON.stringify(mem.title || 'untitled')}`,
449
+ error: `Memory missing required fields (title, type, path, content): ${JSON.stringify(rawMem.title || 'untitled')}`,
336
450
  }));
337
451
  process.exit(1);
338
452
  }
339
453
 
340
- if (!TYPE_DEFAULTS[mem.type]) {
341
- console.error(JSON.stringify({ error: `Unknown memory type: ${mem.type}` }));
454
+ if (!TYPE_DEFAULTS[rawMem.type]) {
455
+ console.error(JSON.stringify({ error: `Unknown memory type: ${rawMem.type}` }));
456
+ process.exit(1);
457
+ }
458
+
459
+ // Provenance gate (ASI06) — decide what this origin is allowed to claim
460
+ // before anything reaches disk.
461
+ const policy = applyOriginPolicy(rawMem);
462
+ if (policy.error) {
463
+ console.error(JSON.stringify({ error: `${policy.error} — memory: ${JSON.stringify(rawMem.title)}` }));
342
464
  process.exit(1);
343
465
  }
466
+ const { origin } = policy;
467
+ const mem = policy.mem;
468
+ for (const c of policy.clamps) clampsReported.push({ ...c, title: rawMem.title });
344
469
 
345
470
  // Generate ID
346
471
  const id = generateId();
347
472
  newIds.push(id);
348
473
 
349
474
  // Compute strength/decay
350
- const { fileContent, strength, decay_rate } = buildMemoryFileContent(mem, id, now);
475
+ const { fileContent, strength, decay_rate } = buildMemoryFileContent(
476
+ mem, id, now, origin, policy.policy.decay_multiplier
477
+ );
351
478
 
352
479
  // Create directories
353
480
  const fullPath = path.join(brainDir, mem.path);
@@ -357,8 +484,31 @@ async function main() {
357
484
  // Write memory file
358
485
  atomicWriteSync(fullPath, fileContent);
359
486
 
487
+ // Record the write before reporting success. An audit failure is surfaced
488
+ // rather than thrown: the memory is already on disk, and silently dropping
489
+ // the trail would be the worse outcome of the two.
490
+ try {
491
+ appendAuditLog(brainDir, {
492
+ ts: now,
493
+ event: 'memorize',
494
+ id,
495
+ origin,
496
+ agent: HOST_AGENT,
497
+ type: mem.type,
498
+ path: mem.path,
499
+ title: mem.title,
500
+ salience: mem.salience ?? 0.5,
501
+ confidence: mem.confidence ?? 0.7,
502
+ decay_rate,
503
+ entrenched: Boolean(mem.pinned || mem.stable),
504
+ clamped: policy.clamps.map((c) => c.field),
505
+ });
506
+ } catch (err) {
507
+ auditErrors.push({ id, error: err.message });
508
+ }
509
+
360
510
  // Update index
361
- const indexEntry = buildIndexEntry(mem, id, strength, decay_rate, now);
511
+ const indexEntry = buildIndexEntry(mem, id, strength, decay_rate, now, origin);
362
512
  addMemory(index, id, indexEntry);
363
513
 
364
514
  // CoALA Phase 1: register a born-pinned memory in the pinned manifest
@@ -413,6 +563,7 @@ async function main() {
413
563
  cognitive_type: mem.cognitive_type || 'semantic',
414
564
  strength,
415
565
  decay_rate,
566
+ origin,
416
567
  salience: mem.salience ?? 0.5,
417
568
  confidence: mem.confidence ?? 0.7,
418
569
  tags: mem.tags || [],
@@ -441,6 +592,10 @@ async function main() {
441
592
  stored: results,
442
593
  total: results.length,
443
594
  index_count: index.memory_count,
595
+ // Only present when the provenance policy actually lowered something, so
596
+ // the agent can tell the user a claim was downgraded and why.
597
+ ...(clampsReported.length ? { provenance_clamps: clampsReported } : {}),
598
+ ...(auditErrors.length ? { audit_errors: auditErrors } : {}),
444
599
  };
445
600
 
446
601
  // Sync if requested
@@ -0,0 +1,125 @@
1
+ ---
2
+ description: Cold-start your brain from transcripts your agents already wrote
3
+ argument-hint: "[--project P] [--since 30d] [--limit N]"
4
+ ---
5
+
6
+ # /brain:import — Cold-Start From Past Sessions
7
+
8
+ A new brain is empty. But your agents have been keeping transcripts for months — decisions, preferences, corrections, conventions. This command reads those transcripts and turns them into real memories.
9
+
10
+ **User input:** $ARGUMENTS
11
+
12
+ **Flags** (passed straight through to the CLI):
13
+ - `--source <id>` — Which history store to read (default `claude-code`; `--sources` lists what's detected)
14
+ - `--project <name>` — Only sessions from one project
15
+ - `--since <30d|6m|2026-01-01>` — Only sessions since then
16
+ - `--limit <N>` — Max sessions to consider (default 40)
17
+ - `--all` — Re-offer sessions already imported
18
+
19
+ ## Your role
20
+
21
+ The CLI harvests. **You distill.** It hands you a digest of past sessions — titles, prompts, projects, files touched, dates. It does not decide what matters. That is the entire job here, and it is a judgement call the CLI deliberately does not make.
22
+
23
+ ## Steps
24
+
25
+ ### 1. Harvest
26
+
27
+ ```bash
28
+ brain import $ARGUMENTS
29
+ ```
30
+
31
+ Returns `{ source, scanned, budget, sessions[] }`. Each session has `session_id`, `title`, `project`, `git_branch`, `started`/`ended`, `turns`, `files_touched`, and a sample of `prompts`.
32
+
33
+ Read `scanned` honestly when you report — the fields mean different things:
34
+ - `skipped` — sessions that held nothing readable (too few real prompts)
35
+ - `filtered` — sessions excluded by the user's own `--project`/`--since` scoping, **not** discarded as junk
36
+ - `already_imported` — sessions a previous run already covered
37
+ - `remaining` — candidates that did not fit this pass; the number to quote when suggesting a re-run
38
+
39
+ If `scanned.files` is 0, tell the user no history was found and stop.
40
+
41
+ ### 2. Check what the brain already knows
42
+
43
+ **Do not skip this on a brain that is already populated.** Import is most often run cold, but it is also run against a brain with hundreds of memories, and there the dominant failure is duplication — re-writing facts that are already stored, which inflates the brain and splits recall across near-identical entries.
44
+
45
+ Before distilling, recall against the themes you can see in the digest:
46
+
47
+ ```bash
48
+ brain recall "<theme from the digest>" --project <project> --top 4
49
+ ```
50
+
51
+ If a theme already scores well against an existing memory, drop it. Only write what is genuinely absent, or what materially *updates* something already stored — and if it's an update, say so in the content rather than filing a near-duplicate.
52
+
53
+ ### 3. Read for durable facts, not events
54
+
55
+ You are looking at what someone *did*, and trying to recover what is still *true*. Those are different things.
56
+
57
+ **Worth remembering** — things that outlive the session:
58
+ - Stated preferences and conventions ("always use X", "I hate Y", "we commit straight to main")
59
+ - Architecture and tooling decisions, with the rationale if it's visible
60
+ - Recurring corrections — if the user pushed back on the same thing three times, that is a preference
61
+ - The shape of each project: what it is, its stack, where it deploys
62
+ - Constraints, deadlines, ongoing goals
63
+
64
+ **Not worth remembering** — things that were only true that afternoon:
65
+ - "fix the failing test", "what does this function do", "run the build"
66
+ - Anything already recorded in the repo, git history, or a CLAUDE.md
67
+ - One-off debugging that ended when the bug was fixed
68
+ - Anything superseded by later work in the digest
69
+
70
+ A good import from 40 sessions is roughly **5–20 memories**, not 40. One memory per session means you transcribed instead of distilled.
71
+
72
+ ### 4. Prefer synthesis across sessions
73
+
74
+ The digest's real value is the pattern across sessions, which no single session contains. Six sessions in the same repo tell you what that project *is* — that is one strong `semantic` memory, worth more than six thin episodic ones.
75
+
76
+ Merge aggressively. Write the memory the user would want to find in six months.
77
+
78
+ ### 5. Classify and write
79
+
80
+ Use the classification rules in `/brain:memorize` — same types, same cognitive types, same hierarchy under `professional/` etc.
81
+
82
+ Two rules specific to import:
83
+
84
+ **`origin` is `agent-inferred`.** You are inferring durable facts from a transcript, not being told them. Do not use `user` — the CLI's provenance gate will reject any attempt to pin or entrench from this path, which is the intended behavior. If a fact deserves entrenchment, the user can `/brain:pin` it deliberately afterwards.
85
+
86
+ **Confidence should reflect the evidence.** A convention stated once in passing is not a convention you are sure about. Something the user repeated across four sessions is. Set `confidence` accordingly rather than defaulting everything to 0.7.
87
+
88
+ Set `encoding_context.project` to the digest's `project` field, and `source` to `"import:<source>:<session_id>"` so the memory traces back to the transcript it came from.
89
+
90
+ Then write them:
91
+
92
+ ```bash
93
+ brain memorize <<'EOF'
94
+ { "memories": [ ... ] }
95
+ EOF
96
+ ```
97
+
98
+ ### 6. Mark what you used
99
+
100
+ ```bash
101
+ brain import --mark <session_id> <session_id> ...
102
+ ```
103
+
104
+ Mark **every session you read**, including the ones you decided held nothing worth keeping — otherwise the next run offers them again and you re-litigate the same judgement. This cursor is what makes import incremental.
105
+
106
+ Use the **full** `session_id` from the digest, never a shortened form you printed for display. The CLI rejects ids that match no real session and exits non-zero; check the `unknown` array if it does.
107
+
108
+ ### 7. Report
109
+
110
+ ```
111
+ ◉ Imported <N> memories from <M> sessions across <P> projects
112
+
113
+ <type> <title>
114
+ ...
115
+
116
+ <scanned.remaining> sessions remain — re-run to continue.
117
+ ```
118
+
119
+ Quote `scanned.remaining` for the re-run line, and mention `scanned.filtered` only if the user passed a scoping flag.
120
+
121
+ ## Notes
122
+
123
+ - **Safe to re-run.** The cursor means a second run offers only new sessions. `--all` overrides it.
124
+ - **Start scoped if the history is large.** `--project <name>` or `--since 30d` gives a better first pass than trying to swallow a year at once.
125
+ - **Nothing is written without you.** `brain import` only ever reads and prints; every memory goes through `brain memorize` and its provenance gate.