memd 0.3.3__tar.gz → 0.4.2__tar.gz

{memd-0.3.3 → memd-0.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memd
-Version: 0.3.3
+Version: 0.4.2
 Summary: Agent-driven project-memory curator
 Author: lowcache
 License-Expression: MIT
@@ -31,8 +31,9 @@ Dynamic: license-file
   <img src="https://raw.githubusercontent.com/lowcache/memd/main/assets/memd.png" alt="memd logo" />
 </p>
 
-[![Release](https://img.shields.io/badge/version-0.3.3-blue.svg)](https://github.com/lowcache/memd)
-[![Python Version Compatibility](https://img.shields.io/badge/python-3.10%20%7C%203.11%20%7C%203.12%20%7C%203.13%20%7C%203.14-brightgreen.svg)](#installation)
+[![CI](https://github.com/lowcache/memd/actions/workflows/ci.yml/badge.svg)](https://github.com/lowcache/memd/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/memd.svg)](https://pypi.org/project/memd/)
+[![Python Versions](https://img.shields.io/pypi/pyversions/memd.svg)](https://pypi.org/project/memd/)
 [![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)](LICENSE)
 [![Nix Flake](https://img.shields.io/badge/Nix-Flake-blueviolet.svg)](#nix-installation)
 
@@ -54,7 +55,7 @@ Every AI coding session ends the same way: the context window fills up, the sess
 
 ## How It Works
 
-Sessions end (or a timer fires), memd collects whatever is new — transcripts, databases, inbox notes — scrubs it for secrets, and hands it to a small "curator" model that rewrites the memory files. Python code, not the model, gets the final say on what's allowed to change.
+Sessions end (or a timer fires), memd collects whatever is new — transcripts, databases, inbox notes — scrubs it for secrets, and hands it to a small "curator" model that applies surgical section-level updates to the memory files. Python code, not the model, gets the final say on what's allowed to change.
 
 ```mermaid
 graph TD
@@ -123,9 +124,9 @@ For two conventions memd doesn't write — GitHub Copilot's `.github/copilot-ins
 
 The core of memd doesn't care which agent or model you use: the memory files are plain markdown, the inbox accepts notes from anything that can write a file, `memd brief` prints to stdout for any tool to consume, and the `curator_cmd` config key lets any LLM command do the distilling. The parts that are currently tool-specific:
 
-*   **Automatic transcript ingestion** understands `claude-code` session logs and `antigravity-cli` databases. Other agents' transcripts can be wired in per-project via `extra_sources` (plain text / JSONL), or their facts routed through the inbox — there is no built-in parser for them yet.
+*   **Automatic transcript ingestion** understands `claude-code` session logs and `antigravity-cli` databases natively, and digests *any other* CLI's session logs via content sniffing and a best-effort text fallback (common `user:`/`assistant:` markers become turns) — point `extra_sources` at whatever a tool writes, or route facts through the inbox instead.
 *   **Lifecycle hooks** (`memd install-hooks`, session-start brief injection) exist only for `claude-code`. Every other tool relies on the periodic sweep timer instead: slower cadence, same result.
-*   **The default curator is the `claude` CLI**, and the default `model_small`/`model_large` names (`haiku`/`sonnet`) assume it. Set `curator_cmd` to distill with a local model, another vendor's CLI, or anything else — [`contrib/curators/`](contrib/curators/) has ready-made wrappers and the backend contract, so you can run memd with no Anthropic dependency at all.
+*   **The default curator is the `claude` CLI**, and the default `model_small`/`model_large` names (both `sonnet`) assume it. Set `curator_cmd` to distill with a local model, another vendor's CLI, or anything else — [`contrib/curators/`](contrib/curators/) has ready-made wrappers and the backend contract, so you can run memd with no Anthropic dependency at all.
 
 ---
 
@@ -181,20 +182,22 @@ For complete writer/reader guidelines, see [INBOX-PROTOCOL.md](INBOX-PROTOCOL.md
 
 Distilling means sending session content to an LLM, so it's worth being precise about what goes where and what it costs.
 
-**What gets sent.** Whatever is new in your transcripts since the last run, capped at `digest_cap_chars` (60,000 by default), plus the current contents of the four memory files. Credentials are scrubbed first — 13 patterns, [detailed below](#built-in-credential-redaction) — but that's shape-matching, not a guarantee.
+**What gets sent.** Whatever is new in your transcripts since the last run — as chronologically ordered, timestamped session spans — capped at `digest_cap_chars` (60,000 by default), plus the current memory files (`mistakes.md` only as a heading index; it is append-only). Credentials are scrubbed first — 13 patterns, [detailed below](#built-in-credential-redaction) — but that's shape-matching, not a guarantee.
 
 **Where it goes.** To whatever `curator_cmd` points at. Out of the box that's the `claude` CLI, which means your session content reaches Anthropic's API. Point it at a local model and nothing leaves the machine; [`contrib/curators/`](contrib/curators/) has a verified ollama wrapper and the backend contract. `memd status` always prints the backend in use, so you never have to infer it:
 
 ```
 $ memd status
-curator      : claude (headless)  [models: haiku/sonnet]  — sends session content to the Anthropic API
+curator      : claude (headless)  [models: sonnet/sonnet]  — sends session content to the Anthropic API
 ```
 
-**What it costs.** A distill runs when a session ends, and the sweep timer checks every 30 minutes for projects with new content. Most sweeps do nothing — a project is skipped unless it has unread transcript content or an inbox note, and transcripts touched in the last `quiet_seconds` (10 minutes) are left alone so an active session isn't digested mid-flight. When a distill does run it's one call, on `model_small` (haiku) unless the digest exceeds `escalate_chars` (15,000), which promotes it to `model_large` (sonnet). To spend less: raise `quiet_seconds`, lengthen the timer interval, or set both models to something cheap. To spend nothing, use a local backend.
+**What it costs.** A distill runs when a session ends, and the sweep timer checks every 30 minutes for projects with new content. Most sweeps do nothing — a project is skipped unless it has unread transcript content or an inbox note, and transcripts touched in the last `quiet_seconds` (10 minutes) are left alone so an active session isn't digested mid-flight. When a distill does run it's one call, on `model_small` unless the digest exceeds `escalate_chars` (15,000), which promotes it to `model_large`.
+
+Both default to `sonnet`, which is a deliberate choice: curation quality *is* the product, and a weak default model is indistinguishable from a weak tool to anyone trying memd for the first time. It costs more than the cheapest option. To spend less, set `model_small` to `haiku` — routine distills drop to it while large backlogs still escalate. To spend nothing, use a local backend.
 
 **What memd never does.** No network calls of its own, no telemetry, no phone-home. It reads session transcripts, inbox notes, and any `extra_sources` you configure — not your source code. It never executes anything from a transcript or from curator output. See [SECURITY.md](SECURITY.md).
 
-**When ingestion breaks, it says so.** memd reads formats that belong to other tools — `claude-code`'s session logs, `antigravity-cli`'s databases — and those can change without notice. Parsing is deliberately forgiving, so a change doesn't crash anything; the risk is the opposite one, that memd keeps running and quietly stops learning. So it watches for exactly that: session content read, nothing usable extracted. `memd status` reports it instead of leaving you to notice months later that memory went stale.
+**When ingestion breaks, it says so.** memd reads formats that belong to other tools — `claude-code`'s session logs, `antigravity-cli`'s databases — and those can change without notice. Parsing is deliberately forgiving, so a change doesn't crash anything; unknown third-party formats simply degrade to the text fallback and keep working. The surviving blindspot is claude-shaped JSONL whose entries stop yielding lines: session content read, nothing usable extracted. `memd status` reports exactly that instead of leaving you to notice months later that memory went stale.
 
 ```
 ! ingestion  : 9720 bytes read since 2026-07-30T06:11:27, nothing usable extracted.
@@ -205,7 +208,7 @@ curator      : claude (headless)  [models: haiku/sonnet]  — sends session cont
 Nothing is lost when this happens: read cursors only advance after a successful distill, so the unread content simply replays once parsing works again.
 
 ### 4. Extra Custom Sources
-Additional text transcript sources (e.g. custom JSONL logs) can be registered per-project under the `projects.<path>.extra_sources` array in `config.json`.
+Additional transcript sources (any text or JSONL log another CLI writes — role tags like `user:`/`assistant:` are detected heuristically) can be registered per-project under the `projects.<path>.extra_sources` array in `config.json`.
 
 ---
 
@@ -220,10 +223,11 @@ Configuration is stored in `$XDG_CONFIG_HOME/memd/config.json` (defaults to `~/.
 | `claude_bin` | `string` | `"claude"` | Name or path of the `claude-code` binary. |
 | `curator_cmd` | `array` | `[]` | Override list of arguments to invoke a custom curator backend (e.g., `["nix", "run", ".#", "--", "status"]`). Substitutes `{model}`. If empty, falls back to the headless `claude_bin` command. |
 | `antigravity_dir` | `string` | `"~/.gemini/antigravity-cli"` | Directory containing native SQLite database conversations. |
-| `model_small` | `string` | `"haiku"` | Model used for small/routine transcript distills. |
-| `model_large` | `string` | `"sonnet"` | Model used for large/complex session distills. |
-| `escalate_chars` | `integer` | `15000` | Character count above which digests are run with `model_large`. |
-| `digest_cap_chars` | `integer` | `60000` | Maximum length of transcript digest fed to the LLM. |
+| `model_small` | `string` | `"sonnet"` | Model used for routine distills. Set to `"haiku"` to cut cost. |
+| `model_large` | `string` | `"sonnet"` | Model used once a digest exceeds `escalate_chars`. |
+| `escalate_chars` | `integer` | `15000` | Digest size above which `model_large` is used, whatever triggered the distill. |
+| `digest_cap_chars` | `integer` | `60000` | Maximum length of transcript digest fed to the LLM. Oversize digests drop whole low-signal sessions (named in an omission marker) rather than truncating mid-story. |
+| `capture_sidechains` | `boolean` | `true` | Include subagent ("sidechain") work from transcripts as indented secondary-signal lines. |
 | `quiet_seconds` | `integer` | `600` | Skip digesting transcripts modified within this cooldown period. |
 | `sweep_jobs` | `integer` | `4` | Number of worker threads for parallel sweeps. |
 | `auto_scaffold` | `boolean` | `true` | Auto-create `.memory/` structure inside detected git repositories. |
@@ -395,7 +399,7 @@ Symlink the shim onto your `PATH` for a live-updating install: `ln -s "$PWD/memd
 
 ## Testing & Verification
 
-180 automated tests cover the parts that would hurt if they broke: XDG isolation, cursors, redaction, database parsing, atomic inbox publishing (including an 8-writer concurrency stress test), retry policies, curator output parsing (including the malformed JSON small models produce), brief budgeting, memory-repo decoupling + worktree sharing, silent-ingestion-breakage detection, and curation quality. Verified on Python 3.10 through 3.14.
+184 automated tests cover the parts that would hurt if they broke: XDG isolation, cursors, redaction, database parsing, atomic inbox publishing (including an 8-writer concurrency stress test), retry policies, curator output parsing (including the malformed JSON small models produce), brief budgeting, memory-repo decoupling + worktree sharing, silent-ingestion-breakage detection, and curation quality. Verified on Python 3.10 through 3.14.
 
 To run tests in a Nix-isolated environment:
 ```bash

{memd-0.3.3 → memd-0.4.2}/README.md

@@ -4,8 +4,9 @@
   <img src="https://raw.githubusercontent.com/lowcache/memd/main/assets/memd.png" alt="memd logo" />
 </p>
 
-[![Release](https://img.shields.io/badge/version-0.3.3-blue.svg)](https://github.com/lowcache/memd)
-[![Python Version Compatibility](https://img.shields.io/badge/python-3.10%20%7C%203.11%20%7C%203.12%20%7C%203.13%20%7C%203.14-brightgreen.svg)](#installation)
+[![CI](https://github.com/lowcache/memd/actions/workflows/ci.yml/badge.svg)](https://github.com/lowcache/memd/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/memd.svg)](https://pypi.org/project/memd/)
+[![Python Versions](https://img.shields.io/pypi/pyversions/memd.svg)](https://pypi.org/project/memd/)
 [![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)](LICENSE)
 [![Nix Flake](https://img.shields.io/badge/Nix-Flake-blueviolet.svg)](#nix-installation)
 
@@ -27,7 +28,7 @@ Every AI coding session ends the same way: the context window fills up, the sess
 
 ## How It Works
 
-Sessions end (or a timer fires), memd collects whatever is new — transcripts, databases, inbox notes — scrubs it for secrets, and hands it to a small "curator" model that rewrites the memory files. Python code, not the model, gets the final say on what's allowed to change.
+Sessions end (or a timer fires), memd collects whatever is new — transcripts, databases, inbox notes — scrubs it for secrets, and hands it to a small "curator" model that applies surgical section-level updates to the memory files. Python code, not the model, gets the final say on what's allowed to change.
 
 ```mermaid
 graph TD
@@ -96,9 +97,9 @@ For two conventions memd doesn't write — GitHub Copilot's `.github/copilot-ins
 
 The core of memd doesn't care which agent or model you use: the memory files are plain markdown, the inbox accepts notes from anything that can write a file, `memd brief` prints to stdout for any tool to consume, and the `curator_cmd` config key lets any LLM command do the distilling. The parts that are currently tool-specific:
 
-*   **Automatic transcript ingestion** understands `claude-code` session logs and `antigravity-cli` databases. Other agents' transcripts can be wired in per-project via `extra_sources` (plain text / JSONL), or their facts routed through the inbox — there is no built-in parser for them yet.
+*   **Automatic transcript ingestion** understands `claude-code` session logs and `antigravity-cli` databases natively, and digests *any other* CLI's session logs via content sniffing and a best-effort text fallback (common `user:`/`assistant:` markers become turns) — point `extra_sources` at whatever a tool writes, or route facts through the inbox instead.
 *   **Lifecycle hooks** (`memd install-hooks`, session-start brief injection) exist only for `claude-code`. Every other tool relies on the periodic sweep timer instead: slower cadence, same result.
-*   **The default curator is the `claude` CLI**, and the default `model_small`/`model_large` names (`haiku`/`sonnet`) assume it. Set `curator_cmd` to distill with a local model, another vendor's CLI, or anything else — [`contrib/curators/`](contrib/curators/) has ready-made wrappers and the backend contract, so you can run memd with no Anthropic dependency at all.
+*   **The default curator is the `claude` CLI**, and the default `model_small`/`model_large` names (both `sonnet`) assume it. Set `curator_cmd` to distill with a local model, another vendor's CLI, or anything else — [`contrib/curators/`](contrib/curators/) has ready-made wrappers and the backend contract, so you can run memd with no Anthropic dependency at all.
 
 ---
 
@@ -154,20 +155,22 @@ For complete writer/reader guidelines, see [INBOX-PROTOCOL.md](INBOX-PROTOCOL.md
 
 Distilling means sending session content to an LLM, so it's worth being precise about what goes where and what it costs.
 
-**What gets sent.** Whatever is new in your transcripts since the last run, capped at `digest_cap_chars` (60,000 by default), plus the current contents of the four memory files. Credentials are scrubbed first — 13 patterns, [detailed below](#built-in-credential-redaction) — but that's shape-matching, not a guarantee.
+**What gets sent.** Whatever is new in your transcripts since the last run — as chronologically ordered, timestamped session spans — capped at `digest_cap_chars` (60,000 by default), plus the current memory files (`mistakes.md` only as a heading index; it is append-only). Credentials are scrubbed first — 13 patterns, [detailed below](#built-in-credential-redaction) — but that's shape-matching, not a guarantee.
 
 **Where it goes.** To whatever `curator_cmd` points at. Out of the box that's the `claude` CLI, which means your session content reaches Anthropic's API. Point it at a local model and nothing leaves the machine; [`contrib/curators/`](contrib/curators/) has a verified ollama wrapper and the backend contract. `memd status` always prints the backend in use, so you never have to infer it:
 
 ```
 $ memd status
-curator      : claude (headless)  [models: haiku/sonnet]  — sends session content to the Anthropic API
+curator      : claude (headless)  [models: sonnet/sonnet]  — sends session content to the Anthropic API
 ```
 
-**What it costs.** A distill runs when a session ends, and the sweep timer checks every 30 minutes for projects with new content. Most sweeps do nothing — a project is skipped unless it has unread transcript content or an inbox note, and transcripts touched in the last `quiet_seconds` (10 minutes) are left alone so an active session isn't digested mid-flight. When a distill does run it's one call, on `model_small` (haiku) unless the digest exceeds `escalate_chars` (15,000), which promotes it to `model_large` (sonnet). To spend less: raise `quiet_seconds`, lengthen the timer interval, or set both models to something cheap. To spend nothing, use a local backend.
+**What it costs.** A distill runs when a session ends, and the sweep timer checks every 30 minutes for projects with new content. Most sweeps do nothing — a project is skipped unless it has unread transcript content or an inbox note, and transcripts touched in the last `quiet_seconds` (10 minutes) are left alone so an active session isn't digested mid-flight. When a distill does run it's one call, on `model_small` unless the digest exceeds `escalate_chars` (15,000), which promotes it to `model_large`.
+
+Both default to `sonnet`, which is a deliberate choice: curation quality *is* the product, and a weak default model is indistinguishable from a weak tool to anyone trying memd for the first time. It costs more than the cheapest option. To spend less, set `model_small` to `haiku` — routine distills drop to it while large backlogs still escalate. To spend nothing, use a local backend.
 
 **What memd never does.** No network calls of its own, no telemetry, no phone-home. It reads session transcripts, inbox notes, and any `extra_sources` you configure — not your source code. It never executes anything from a transcript or from curator output. See [SECURITY.md](SECURITY.md).
 
-**When ingestion breaks, it says so.** memd reads formats that belong to other tools — `claude-code`'s session logs, `antigravity-cli`'s databases — and those can change without notice. Parsing is deliberately forgiving, so a change doesn't crash anything; the risk is the opposite one, that memd keeps running and quietly stops learning. So it watches for exactly that: session content read, nothing usable extracted. `memd status` reports it instead of leaving you to notice months later that memory went stale.
+**When ingestion breaks, it says so.** memd reads formats that belong to other tools — `claude-code`'s session logs, `antigravity-cli`'s databases — and those can change without notice. Parsing is deliberately forgiving, so a change doesn't crash anything; unknown third-party formats simply degrade to the text fallback and keep working. The surviving blindspot is claude-shaped JSONL whose entries stop yielding lines: session content read, nothing usable extracted. `memd status` reports exactly that instead of leaving you to notice months later that memory went stale.
 
 ```
 ! ingestion  : 9720 bytes read since 2026-07-30T06:11:27, nothing usable extracted.
@@ -178,7 +181,7 @@ curator      : claude (headless)  [models: haiku/sonnet]  — sends session cont
 Nothing is lost when this happens: read cursors only advance after a successful distill, so the unread content simply replays once parsing works again.
 
 ### 4. Extra Custom Sources
-Additional text transcript sources (e.g. custom JSONL logs) can be registered per-project under the `projects.<path>.extra_sources` array in `config.json`.
+Additional transcript sources (any text or JSONL log another CLI writes — role tags like `user:`/`assistant:` are detected heuristically) can be registered per-project under the `projects.<path>.extra_sources` array in `config.json`.
 
 ---
 
@@ -193,10 +196,11 @@ Configuration is stored in `$XDG_CONFIG_HOME/memd/config.json` (defaults to `~/.
 | `claude_bin` | `string` | `"claude"` | Name or path of the `claude-code` binary. |
 | `curator_cmd` | `array` | `[]` | Override list of arguments to invoke a custom curator backend (e.g., `["nix", "run", ".#", "--", "status"]`). Substitutes `{model}`. If empty, falls back to the headless `claude_bin` command. |
 | `antigravity_dir` | `string` | `"~/.gemini/antigravity-cli"` | Directory containing native SQLite database conversations. |
-| `model_small` | `string` | `"haiku"` | Model used for small/routine transcript distills. |
-| `model_large` | `string` | `"sonnet"` | Model used for large/complex session distills. |
-| `escalate_chars` | `integer` | `15000` | Character count above which digests are run with `model_large`. |
-| `digest_cap_chars` | `integer` | `60000` | Maximum length of transcript digest fed to the LLM. |
+| `model_small` | `string` | `"sonnet"` | Model used for routine distills. Set to `"haiku"` to cut cost. |
+| `model_large` | `string` | `"sonnet"` | Model used once a digest exceeds `escalate_chars`. |
+| `escalate_chars` | `integer` | `15000` | Digest size above which `model_large` is used, whatever triggered the distill. |
+| `digest_cap_chars` | `integer` | `60000` | Maximum length of transcript digest fed to the LLM. Oversize digests drop whole low-signal sessions (named in an omission marker) rather than truncating mid-story. |
+| `capture_sidechains` | `boolean` | `true` | Include subagent ("sidechain") work from transcripts as indented secondary-signal lines. |
 | `quiet_seconds` | `integer` | `600` | Skip digesting transcripts modified within this cooldown period. |
 | `sweep_jobs` | `integer` | `4` | Number of worker threads for parallel sweeps. |
 | `auto_scaffold` | `boolean` | `true` | Auto-create `.memory/` structure inside detected git repositories. |
@@ -368,7 +372,7 @@ Symlink the shim onto your `PATH` for a live-updating install: `ln -s "$PWD/memd
 
 ## Testing & Verification
 
-180 automated tests cover the parts that would hurt if they broke: XDG isolation, cursors, redaction, database parsing, atomic inbox publishing (including an 8-writer concurrency stress test), retry policies, curator output parsing (including the malformed JSON small models produce), brief budgeting, memory-repo decoupling + worktree sharing, silent-ingestion-breakage detection, and curation quality. Verified on Python 3.10 through 3.14.
+184 automated tests cover the parts that would hurt if they broke: XDG isolation, cursors, redaction, database parsing, atomic inbox publishing (including an 8-writer concurrency stress test), retry policies, curator output parsing (including the malformed JSON small models produce), brief budgeting, memory-repo decoupling + worktree sharing, silent-ingestion-breakage detection, and curation quality. Verified on Python 3.10 through 3.14.
 
 To run tests in a Nix-isolated environment:
 ```bash

{memd-0.3.3 → memd-0.4.2}/memd/__init__.py

@@ -15,7 +15,7 @@ the invariants the model must not be trusted with: frontmatter, append-only
 mistakes.md, shrink guard, size budgets, archive overflow, git audit commits.
 """
 
-__version__ = "0.3.3"
+__version__ = "0.4.2"
 
 from memd.errors import ConfigError, CuratorError, DigestError, MemdError
 from memd.config import (
@@ -87,11 +87,17 @@ from memd.digest import (
     digest_ag_db,
     digest_jsonl,
     digest_source,
+    digest_text_log,
     project_ag_dbs,
     redact,
+    render_span,
+    score_span,
+    sniff_and_digest,
+    sort_spans,
+    truncate_spans,
 )
 from memd.inbox import collect_inbox, write_inbox_note
-from memd.curator import CHARTER, build_prompt, call_curator, validate
+from memd.curator import CHARTER, apply_ops, build_prompt, call_curator, validate
 from memd.sync import (
     INGEST_STALL_BYTES,
     encode_claude_dir,

{memd-0.3.3 → memd-0.4.2}/memd/cli.py

@@ -54,6 +54,14 @@ def cmd_status(cfg):
         print(f"  last distill : {m.get('last_sync', 'never')}{detail}")
         if m.get("summary"):
             print(f"  summary      : {m['summary']}")
+        # rolling journal: recent distill events make gaps visible (a project
+        # with pending sources but only failed outcomes is quietly blind)
+        hist = [h for h in m.get("history", []) if isinstance(h, dict)]
+        for h in hist[-2:]:
+            outcome = h.get("outcome", "?")
+            mark = "!" if outcome == "fail" else " "
+            print(f"  {mark} {h.get('ts', '?')} [{h.get('trigger', '?')}/"
+                  f"{h.get('model', '?')}] {str(h.get('summary', ''))[:90]}")
         print(f"  sources      : {len(srcs)} ({ag} antigravity), "
               f"{len(pending)} with pending content, {inbox} inbox note(s)")
         stalled = m.get("ingest_stalled_bytes")

{memd-0.3.3 → memd-0.4.2}/memd/config.py

@@ -31,10 +31,19 @@ DEFAULT_CONFIG = {
                                     # contain one JSON object (fences/prose ok).
                                     # Empty -> claude_bin headless invocation.
     "antigravity_dir": os.path.join(HOME, ".gemini", "antigravity-cli"),
-    "model_small": "haiku",
+    # Both tiers default to the same capable model, deliberately. Curation is
+    # the product, and someone evaluating memd for the first time cannot tell
+    # "this tool is mediocre" from "this tool's default model is mediocre" —
+    # they just stop using it, having never found the config key. Good distills
+    # out of the box are worth more than the token savings. Set model_small to
+    # something cheaper (e.g. "haiku") once you trust it.
+    "model_small": "sonnet",
     "model_large": "sonnet",
-    "escalate_chars": 15000,        # session-end digests above this go to model_large
+    "escalate_chars": 15000,        # digests above this go to model_large
     "digest_cap_chars": 60000,      # max transcript digest fed to the model
+    # subagent (sidechain) turns ride along in digests as indented "  SUB> "
+    # lines — delegated work reaches memory instead of vanishing
+    "capture_sidechains": True,
     "quiet_seconds": 600,           # sweep skips transcripts modified more recently
     "sweep_jobs": 4,                # parallel project syncs per sweep (threads;
                                     # per-project flock still serializes any