File size: 15,526 Bytes
817dc1b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 | {
"model": "CludeMem-E4B",
"about": "One entry per memory operation. Send the task's system_prompt as the system message and the input, formatted as user_format describes, as the user message; the reply is one JSON object with output_keys. output_values lists the values those fields take in the training targets. The examples are test items whose inputs do not occur in the training data; example.output is the reference reply.",
"chat_template": {
"apply_chat_template_kwargs": {
"add_generation_prompt": true,
"enable_thinking": true
},
"rendered_form": "<bos><|turn>system\n<|think|>\n{system}<turn|>\n<|turn>user\n{user}<turn|>\n<|turn>model\n",
"note": "enable_thinking=True reproduces the training prompt: a '<|think|>' line opens the system turn. The model does not emit a thinking block; it replies with one JSON object. transformers' apply_chat_template omits the line unless enable_thinking=True is passed.",
"tokenizers_checked": [
"google/gemma-4-e4b-it",
"unsloth/gemma-4-E4B-it"
]
},
"decoding": "greedy (temperature 0)",
"tasks": {
"CLASSIFY": {
"system_prompt": "Classify the memory. Output JSON: {type, importance (0-1), tags[], concepts[], emotional_valence (-1..1)}.",
"user_format": {
"template": "{memory text}",
"description": "One memory statement as free text on a single line."
},
"output_keys": [
"type",
"importance",
"tags",
"concepts",
"emotional_valence"
],
"output_values": {
"type": [
"episodic",
"semantic"
],
"importance": [
0.3,
0.4,
0.8,
0.6,
0.5,
0.85,
0.9
],
"emotional_valence": [
0.2,
0,
-0.3,
0.3,
0.5,
-0.4
]
},
"example": {
"system": "Classify the memory. Output JSON: {type, importance (0-1), tags[], concepts[], emotional_valence (-1..1)}.",
"user": "By the way, around then on 2026-02-16, Priya spent the afternoon reviewing the Polaris roadmap, triaged the mobile backlog, and noted follow-ups for the data sync. It came up while we were chatting.",
"output": {
"type": "episodic",
"importance": 0.3,
"tags": [
"event",
"task",
"priya"
],
"concepts": [
"event",
"task"
],
"emotional_valence": 0.2
}
}
},
"EXTRACT": {
"system_prompt": "Extract atomic memories from the text. Output JSON: {memories:[{content, summary, type}]}.",
"user_format": {
"template": "{Speaker}: {utterance}\\n{Speaker}: {utterance}\\n...",
"description": "A dialogue block, one 'Name: utterance' line per turn. Some rows mix in non-dialogue lines (log lines, markdown tables, tool-call JSON, code fences) that carry no memory."
},
"output_keys": [
"memories"
],
"output_values": {
"memories[].type": [
"episodic",
"semantic"
]
},
"example": {
"system": "Extract atomic memories from the text. Output JSON: {memories:[{content, summary, type}]}.",
"user": "Tomas: Actually, scratch what I said before — tomas now prefers matcha, having switched from cola.\nTomas: By the way, around then tomas started a new job at Brightwell on 2026-07-18. It came up while we were chatting.\n{\"tool\":\"search\",\"status\":\"ok\",\"hits\":1}",
"output": {
"memories": [
{
"content": "Tomas now prefers matcha, having switched from cola.",
"summary": "Tomas now prefers matcha, having switched from cola.",
"type": "semantic"
},
{
"content": "Tomas started a new job at Brightwell on 2026-07-18.",
"summary": "Tomas started a new job at Brightwell on 2026-07-18.",
"type": "episodic"
}
]
}
}
},
"ENTITIES": {
"system_prompt": "Extract entities and relations. Output JSON: {entities:[{name,type,aliases}], relations:[{head,type,tail}]}.",
"user_format": {
"template": "{memory text}",
"description": "One memory statement as free text on a single line."
},
"output_keys": [
"entities",
"relations"
],
"output_values": {
"entities[].type": [
"person",
"location",
"organization",
"project"
],
"relations[].type": [
"lives_in",
"works_at",
"leads"
]
},
"example": {
"system": "Extract entities and relations. Output JSON: {entities:[{name,type,aliases}], relations:[{head,type,tail}]}.",
"user": "Actually, scratch what I said before — on 2026-05-02, Theo spent the afternoon reviewing the Beacon roadmap, triaged the mobile backlog, and noted follow-ups for the product sync.",
"output": {
"entities": [
{
"name": "Theo",
"type": "person",
"aliases": []
}
],
"relations": []
}
}
},
"TEMPORAL": {
"system_prompt": "Extract the event date and its temporal links to the known events. Output JSON: {event_date, precision, links:[{type,target}]}.",
"user_format": {
"template": "Today is {YYYY-MM-DD}.\\nNew event [fN]: {text}\\nKnown events:\\n[fN] {text} ({YYYY-MM-DD})\\n...",
"description": "Reference date, the new event with its id, then the known events, each followed by its date in parentheses. A few rows have no known events (the 'Known events:' block is absent)."
},
"output_keys": [
"event_date",
"precision",
"links"
],
"output_values": {
"precision": [
"day"
],
"links[].type": [
"happens_before",
"happens_after"
]
},
"example": {
"system": "Extract the event date and its temporal links to the known events. Output JSON: {event_date, precision, links:[{type,target}]}.",
"user": "Today is 2026-06-01.\nNew event [f57]: I think on 2026-03-19, Nina spent the afternoon reviewing the Quartz roadmap, triaged the mobile backlog, and noted follow-ups for the data sync. At least that's my sense of it.\nKnown events:\n[f25] On 2026-03-01, Nina spent the afternoon reviewing the Beacon roadmap, triaged the infra backlog, and noted follow-ups for the data sync. (2026-03-01)\n[f41] On 2026-03-01, Nina spent the afternoon reviewing the Quartz roadmap, triaged the growth backlog, and noted follow-ups for the data sync. (2026-03-01)\n[f23] On 2026-03-10, Nina spent the afternoon reviewing the Sable roadmap, triaged the growth backlog, and noted follow-ups for the design sync. (2026-03-10)\n[f22] On 2026-03-24, Nina spent the afternoon reviewing the Quartz roadmap, triaged the mobile backlog, and noted follow-ups for the platform sync. (2026-03-24)\n[f14] On 2026-04-04, Nina spent the afternoon reviewing the Beacon roadmap, triaged the mobile backlog, and noted follow-ups for the ops sync. (2026-04-04)\n[f34] On 2026-04-06, Nina spent the afternoon reviewing the Echo roadmap, triaged the infra backlog, and noted follow-ups for the design sync. (2026-04-06)",
"output": {
"event_date": "2026-03-19",
"precision": "day",
"links": [
{
"type": "happens_after",
"target": "f25"
},
{
"type": "happens_after",
"target": "f41"
},
{
"type": "happens_after",
"target": "f23"
},
{
"type": "happens_before",
"target": "f22"
},
{
"type": "happens_before",
"target": "f14"
},
{
"type": "happens_before",
"target": "f34"
}
]
}
}
},
"CONSOLIDATE": {
"system_prompt": "Consolidate the memories into evidence-linked insights. Output JSON: {insights:[{content, evidence[]}]}.",
"user_format": {
"template": "[fN] {memory}\\n[fN] {memory}\\n...",
"description": "One memory per line, each prefixed with its id."
},
"output_keys": [
"insights"
],
"output_values": {},
"example": {
"system": "Consolidate the memories into evidence-linked insights. Output JSON: {insights:[{content, evidence[]}]}.",
"user": "[f1] Kai lives in Austin.\n[f2] Kai moved to Seoul on 2026-10-21.\n[f3] Kai currently resides in Seoul.\n[f4] Kai enjoys living in Seoul.\n[f5] Kai prefers cola over matcha.\n[f6] Kai now prefers matcha, having switched from cola.\n[f7] Kai started a new job at Aperture on 2026-08-16.\n[f8] Kai is allergic to gluten.",
"output": {
"insights": [
{
"content": "Kai lives in Austin.",
"evidence": [
"f1"
]
},
{
"content": "Kai currently resides in Seoul.",
"evidence": [
"f3"
]
},
{
"content": "Kai enjoys living in Seoul.",
"evidence": [
"f4"
]
},
{
"content": "Kai prefers cola over matcha.",
"evidence": [
"f5"
]
},
{
"content": "Kai now prefers matcha, having switched from cola.",
"evidence": [
"f6"
]
},
{
"content": "Kai is allergic to gluten.",
"evidence": [
"f8"
]
}
]
}
}
},
"COMPACT": {
"system_prompt": "Compact the old memories into one summary, preserving entities and the date range. Output JSON: {summary, preserved_entities[], date_range:{start,end}}.",
"user_format": {
"template": "[fN] {memory}\\n[fN] {memory}\\n...",
"description": "One old memory per line, each prefixed with its id."
},
"output_keys": [
"summary",
"preserved_entities",
"date_range"
],
"output_values": {},
"example": {
"system": "Compact the old memories into one summary, preserving entities and the date range. Output JSON: {summary, preserved_entities[], date_range:{start,end}}.",
"user": "[f2] Aisha moved to Lisbon on 2026-03-18.\n[f7] Aisha started a new job at Northstar on 2026-10-22.\n[f9] The Beacon project shipped on 2026-04-13.",
"output": {
"summary": "Aisha: 3 events covering Aisha, Lisbon, Northstar, Beacon.",
"preserved_entities": [
"Aisha",
"Lisbon",
"Northstar",
"Beacon"
],
"date_range": {
"start": "2026-03-18",
"end": "2026-10-22"
}
}
}
},
"RECONCILE": {
"system_prompt": "Decide how the two memories relate. Output JSON: {verdict, resolution, weaker_id, confidence}.",
"user_format": {
"template": "Memory A [fN]: {text}\\nMemory B [fN]: {text}",
"description": "Exactly two lines, the two memories to compare with their ids. weaker_id in the output names one of these ids, or is null."
},
"output_keys": [
"verdict",
"resolution",
"weaker_id",
"confidence"
],
"output_values": {
"verdict": [
"contradicts",
"supersedes",
"duplicate",
"consistent"
],
"confidence": [
0.9,
0.8,
0.75,
0.95,
0.85
],
"resolution": [
"Memory A and Memory B cannot be ordered; the conflict is unresolved.",
"Memory B conflicts with Memory A; treat B as current.",
"Memory B is the current truth; Memory A is outdated.",
"Memory B restates Memory A; keep one.",
"Memory A and Memory B agree; both are kept.",
"Separate periods; both memories are kept.",
"Different attributes; both memories are kept.",
"Memory A is dated later than Memory B; treat Memory A as current.",
"Memory B is dated later than Memory A; treat Memory B as current.",
"Different people; both memories are kept.",
"The memories overlap in time and cannot be ordered; the conflict is unresolved.",
"Memory B restates Memory A with a coarser date; keep Memory A.",
"Memory A is outdated; Memory B is the current truth.",
"Memory A restates Memory B with a coarser date; keep Memory B.",
"Memory B restates Memory A using another form of the same value; keep Memory A.",
"Memory B is outdated; Memory A is the current truth.",
"Memory A restates Memory B using another form of the same value; keep Memory B.",
"Memory B names the same person as Memory A by a short form; keep Memory A.",
"Memory A names the same person as Memory B by a short form; keep Memory B."
],
"weaker_id": [
"one of the two input ids",
"null"
]
},
"example": {
"system": "Decide how the two memories relate. Output JSON: {verdict, resolution, weaker_id, confidence}.",
"user": "Memory A [f2]: Caleb moved to Toronto on 2026-03-06.\nMemory B [f4]: Caleb enjoys living in Toronto.",
"output": {
"verdict": "consistent",
"resolution": "Memory A and Memory B agree; both are kept.",
"weaker_id": null,
"confidence": 0.9
}
}
},
"QUERY": {
"system_prompt": "Understand the query. Output JSON: {expanded_queries[], temporal_constraints, type_filters[], entities[], intent}.",
"user_format": {
"template": "Today is {YYYY-MM-DD}.\\n{question}",
"description": "Reference date on the first line, the user's question on the second."
},
"output_keys": [
"expanded_queries",
"temporal_constraints",
"type_filters",
"entities",
"intent"
],
"output_values": {
"intent": [
"lookup",
"temporal_lookup"
],
"type_filters": [
[]
]
},
"example": {
"system": "Understand the query. Output JSON: {expanded_queries[], temporal_constraints, type_filters[], entities[], intent}.",
"user": "Today is 2026-06-01.\nDid Zara move to Lyon on 2026-04-19?",
"output": {
"expanded_queries": [
"Did Zara move to Lyon on 2026-04-19?",
"Did Zara move to Lyon on 2026-04-19"
],
"temporal_constraints": {
"after": null,
"before": null
},
"type_filters": [],
"entities": [
"Zara",
"Lyon"
],
"intent": "temporal_lookup"
}
}
},
"ANSWER": {
"system_prompt": "Answer ONLY from the provided memories; abstain if unsupported. Output JSON: {rationale, answer, citations[], confidence, abstain}.",
"user_format": {
"template": "Question: {question}\\n\\nMemories:\\n[fN] {memory}\\n[fN] {memory}\\n...",
"description": "The question, a blank line, 'Memories:', then one retrieved memory per line with its id. citations in the output are ids from this list."
},
"output_keys": [
"rationale",
"answer",
"citations",
"confidence",
"abstain"
],
"output_values": {
"abstain": [
false,
true
],
"confidence": [
0.9,
0.8
],
"answer (when abstain is true)": [
"I do not have enough information to answer that."
]
},
"example": {
"system": "Answer ONLY from the provided memories; abstain if unsupported. Output JSON: {rationale, answer, citations[], confidence, abstain}.",
"user": "Question: Did Tomas move to Lisbon on 2026-03-14?\n\nMemories:\n[f1] Tomas lives in Toronto.\n[f2] Tomas moved to Lisbon on 2026-03-12.",
"output": {
"rationale": "No provided memory answers this.",
"answer": "I do not have enough information to answer that.",
"citations": [],
"confidence": 0.8,
"abstain": true
}
},
"example_answerable": {
"system": "Answer ONLY from the provided memories; abstain if unsupported. Output JSON: {rationale, answer, citations[], confidence, abstain}.",
"user": "Question: Where does Maya live now?\n\nMemories:\n[f2] Maya moved to Austin on 2026-10-22.\n[f1] Maya lives in Lisbon.\n[f3] Maya currently resides in Austin.",
"output": {
"rationale": "Supported by f2.",
"answer": "Austin",
"citations": [
"f2"
],
"confidence": 0.9,
"abstain": false
}
}
}
}
}
|