Model profiles

Every integration resolves a profile from the running model id, then tunes memory behavior per family. Profiles are data in packages/memex-core/lib/model-profiles.js; no host dependency is involved.

Family profile	Match examples	Window	Est. chars/token	Injection	Catalog / recall budget	Digest style
`claude`	`claude-sonnet-4`, `claude-opus-4`	200K (1M beta)	3.6	XML tags	1400 / 600	bullets + `ids:`
`openai`	`gpt-5`, `gpt-5-codex`, `o3`	400K (1M variants)	4.0	Markdown	1200 / 500	numbered + `ids:`
`gemini`	`gemini-2.5-pro`, `gemini-2.0-flash`	1M (2M variants)	4.0	Markdown	1600 / 600	table + `ids:`
`deepseek`	`deepseek-chat`, `deepseek-reasoner`	160K	3.4	Markdown	900 / 400	bilingual + `ids:`
`qwen`	`qwen-max`, `qwen3-coder`	128K–1M	3.5	Markdown	900 / 400	bilingual + `ids:`
`kimi`	`kimi-k2`, `moonshot-v1`	128K–256K	3.5	Markdown	900 / 400	bilingual + `ids:`
`glm`	`glm-4.6`, `chatglm`	128K–200K	3.6	Markdown	900 / 400	bilingual + `ids:`
`llama`	`llama-3.1`, `mistral`, `mixtral`	128K–1M	4.1	Plain text	800 / 350	numbered + `ids:`
`xai`	`grok-4`, `grok-3`	256K–2M	3.8	Markdown	1000 / 450	numbered + `ids:`
`cohere`	`command-r-plus`, `command-a`	128K	4.0	Markdown	900 / 400	numbered + `ids:`
`gemma`	`gemma-3-27b`	128K	4.0	Markdown	800 / 350	numbered + `ids:`
`phi`	`phi-4`, `phi-4-mini`	128K	4.0	Markdown	800 / 350	numbered + `ids:`
`generic`	anything else	128K	4.0	Markdown	900 / 400	numbered + `ids:`

Dedicated current-model variants (v0.3.2 → v0.3.3)

These eight are first-class profiles registered before their family base, so they win resolution while older ids keep the family defaults. Specs verified 2026-10-29 from vendor docs (OpenAI model pages, Anthropic models overview, DeepSeek pricing):

Variant	Model ids	Context	Chars/token	Injection	Catalog / recall	Reasoning mode	Why it differs
`gpt-6.1`	`gpt-6.1-sol`, `gpt-6.1` (+ `-astra`/`sol` ids)	1,050,000 (922k max input; Codex default 500k)	4.0	Markdown	1800 / 700	effort low→max; MCP tooling via Responses API	One consolidated search per step, bigger evidence budget, tags/recency weighted
`claude-opus-5-5`	`claude-opus-5-5` / `claude-opus-5.5`	1,000,000 (128K out; 300K batch)	2.6	XML tags	2000 / 800	adaptive thinking always on, default medium	current tokenizer ≈2.5M chars/1M tokens; decision-oriented XML injection, useWhen boost, archive-first compaction
`claude-sonnet-5-5`	`claude-sonnet-5-5` / `claude-sonnet-5.5`	1,000,000 (128K out; 300K batch)	2.6	XML tags	1600 / 650	adaptive thinking, default high	speed/intelligence balance; smaller budgets than Opus, still citation-first
`deepseek-flash`	`deepseek-flash`, legacy `deepseek-v4-flash`	1,048,576 (384K out)	3.0	Markdown	1200 / 500	thinking / non-thinking (thinking default), 2500 concurrency	latency-sensitive and cheap; bilingual manual, compact structured digests, body/cold-abstract boosts
`gpt-6.1-terra`	`gpt-6.1-terra`	1,050,000	4.0	Markdown	1200 / 500	effort none→max (default medium)	fast/cheap visual+writing work; shallow memory calls, latency-first
`gpt-6-astra`	`gpt-6-astra`, `gpt-6.1-astra`	1,050,000 (922k in / 128K out)	4.0	Markdown	2200 / 900	effort low→max	most capable/expensive; one decisive search/open, compact decision-oriented context
`claude-fable-5-1`	`claude-fable-5-1`	1,000,000 (128K out)	2.6	XML tags	2200 / 900	adaptive thinking always on, default high	demanding reasoning / long-horizon agentic; archive-first, exact ids, body 5.2
`claude-haiku-5-5`	`claude-haiku-5-5`	1,000,000 (128K out)	2.6	XML tags	900 / 400	adaptive thinking, default medium	high-volume latency-sensitive; terse injections, one search/open per step

Sources: GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra, Claude models overview (Opus 5.5 / Sonnet 5.5 / Fable 5.1 / Haiku 5.5), DeepSeek models & pricing.

The variants are checked in this order before the base families, so claude-opus-5-5 never silently falls back to the generic claude profile. Cold-start + long-horizon validation for all eight: profile validation. Overrides still apply on top (resolveModelProfile('gpt-6.1-sol', { windowTokens: 500_000 }) for the Codex default context, for example).

Adapter defaults (written by `init`)

The CLIs write a recommended dedicated profile by default; the integration can override with --model or MEMEX_MODEL:

Adapter	Default `--model`	Recommended alternatives (printed by `init`/`help`)
Claude Code (`memex-claude init`)	`claude-sonnet-5-5`	`claude-opus-5-5`, `claude-fable-5-1`, `claude-haiku-5-5`
Codex (`memex-codex init`)	`gpt-6.1-sol`	`gpt-6.1-terra` (fast/cheap), `gpt-6-astra` (most capable)
MCP server (`memex-mcp`)	unset → `generic`	`memex-mcp --list-profiles` → `deepseek-flash`, `claude-sonnet-5-5`, `gpt-6.1-sol`, `claude-opus-5-5`

init prints the resolved profile (Model profile: claude-sonnet-5-5) and the recommended list, so the written configuration is auditable.

What a profile changes

Budgets — how many tokens of catalog and auto-recall are injected, and how much body memory_open returns by default.
Injection format — XML tags for Claude (<memory-catalog>), Markdown headings for GPT/Gemini/Qwen, plain bracketed blocks for smaller open models (less instruction drift).
Manual style — explicit tool choreography for Claude, terse numbered steps for GPT, table-first reading for Gemini, bilingual (English + Chinese) guidance for DeepSeek/Qwen/Kimi/GLM where CJK prompts are common.
Digest prompt — the compaction-summary prompt used by digest flows, tailored per family, always requiring an ids: line so evicted facts stay retrievable.
Defaults — e.g. Claude/DSH keep both L2+L3, Codex/Gemini default to L3 project memory.

Retrieval reranking

Each profile also carries rerank weights for the hybrid search: the store merges them over DEFAULT_RERANK_WEIGHTS (title 3.0, tags 1.6, summary 1.4, useWhen 1.2, abstract 0.9, body 4.5, coldAnchor 30, recency 0.4, coldPenalty 0.1, bm25 0.8). The body weight applies to body-rescue hits (identifier-like terms found only in a record body). This is the ranking half of model-specific adaptation; the fields above tune injection and prompting.

Profile	Rerank override	Why
`claude`	title 3.2 · useWhen 1.6 · recency 0.3	tool-oriented; “when to read” matters more than freshness
`openai`	tags 1.8 · summary 1.5 · recency 0.55	codex-style agents surface fresh labels and concise summaries
`gemini`	tags 2.0 · summary 1.6 · title 2.8	large-context models tolerate broader recall; tag/summary anchors help
`deepseek`	abstract 1.3 · summary 1.5 · recency 0.35	CJK sessions rely on bounded cold abstracts
`qwen` / `kimi` / `glm`	abstract 1.2 · summary 1.4 · recency 0.3	CJK + long-context; damp recency
`llama`	title 3.4 · abstract 0.7 · recency 0.25	lean ranking, less field noise
`generic`	defaults	fallback

Integrations pass profile.rerank into store.search automatically (MCP, Claude Code hooks, Codex), and profile.queryTerms.maxTerms bounds query extraction (16–24 terms). Override per integration:

new AgentMemory({ model: 'claude-sonnet-4', profileOverrides: { rerank: { tags: 2.2 }, windowTokens: 128_000 } })

Override

# MCP or adapters
node packages/memex-mcp/bin/server.mjs --model my-custom-model --levels l2,l3

# programmatically
import { resolveModelProfile } from '@qiuxu/memex-core/model-profiles'
const profile = resolveModelProfile('my-model', { windowTokens: 64_000, catalogTokens: 500 })

Limits (be honest)

Window sizes and chars/token are conservative estimates, not vendor-tokenizer claims; exact token counts vary per provider and version.
Reranking weights in the store are shared; profiles tune the injection/prompting surface, which is where models differ most in practice.
A profile does not compensate for retrieval misses — combine it with memory_grep/memory_index when search ranks late.
New families are cheap to add: one data entry + a test case; do not fork the core.