feat(recall): cap injected memory context by YOMXXX · Pull Request #71 · Tencent/TencentDB-Agent-Memory

YOMXXX · 2026-05-21T14:36:35Z

Summary

Add recall.maxCharsPerMemory and recall.maxTotalRecallChars with defaults of 0, which do not alter existing behavior. Users can opt in by setting positive values to cap injected memory context.
Apply the budget after L1 search and before <relevant-memories> injection, preserving score order while truncating oversized entries and dropping overflow.
Document the new guards in README, README_CN, and openclaw.plugin.json.

Test Plan

npm test
npm run build
npm pack --dry-run --json
git diff --check

Notes

Addresses L1 auto-recall needs atomization and token-budget safeguards #70 as an MVP focused on recall-time safeguards.
npx tsc --noEmit ... still fails on current upstream/main due unrelated existing type issues in commander/OpenAI settings/offload strict initialization; this PR does not touch those areas.

Signed-off-by: 李冠辰 <liguanchen@xiaomi.com>

Maxwell-Code07 · 2026-05-22T14:12:21Z

感谢您如此及时地解决了 #70 的问题！我们会内部评估，有结论后会尽快反馈。

- Move preExtractMemories to newMessages only (after background/new split) to prevent extracting memories from background context that should only serve as conversational context for the LLM - Remove MEDIUM-confidence hints logging (hints not wired to LLM prompt; keeping types as interface for follow-up PR) - Remove src/ from package.json files field to fix Size Guard limit (matches pattern from Tencent#76 and Tencent#71) - Export callLlmExtraction and passesConfidenceCheck for testability - Add pre-extractor.test.ts covering: - Background messages not pre-extracted - HIGH-confidence dedup via mergeExtractedMemories - Malformed JSON triggers exactly one retry - Confidence filtering does not reject valid persona/instruction

withRiver · 2026-05-25T09:34:40Z

DO NOT modify this file, which might cause some problems. We have increased the limitation of Size Guard from 512KB to 2048KB.

Resolved in 8031ec0: restored package.json to match upstream/main and removed the manifest / extension entry changes from this PR. Fresh checks are green, including Size Guard under the updated 2048KB limit.

YOMXXX · 2026-05-25T09:54:48Z

Thanks for the review. I restored package.json to match upstream/main and removed the package manifest / extension entry changes from this PR.

Fresh local verification after the update:

npm test ✅
npm run build ✅
npm pack --dry-run ✅ package size 647.5 KB, below the updated 2048 KB Size Guard limit.

withRiver · 2026-05-26T07:06:11Z

    recall: {
      enabled: bool(recallGroup, "enabled") ?? true,
      maxResults: num(recallGroup, "maxResults") ?? 5,
+      maxCharsPerMemory: normalizeNonNegativeInt(num(recallGroup, "maxCharsPerMemory"), 800),


normalizeNonNegativeInt() is unnecessary here.

All other config fields (maxResults, scoreThreshold, timeoutMs) use the simple num(...) ?? defaultValue pattern — these two fields don't need special treatment.

Defensive handling for negative/non-integer values is already done on the consumer side in auto-recall.ts's normalizeBudgetLimit() (<= 0 → treated as disabled, Math.floor() for rounding). No need to duplicate that at the parse layer.

The defaults should be 0 (disabled), not 800 / 3000.

This is a new feature — the default behavior should remain backward-compatible. Existing users upgrading should not suddenly see their recall results being truncated or dropped. Those who need this guard can opt in by explicitly setting the values.

Resolved in c588558: recall budget config now follows the existing num(...) ?? 0 pattern, defaults are disabled, and normalizeNonNegativeInt() was removed.

withRiver · 2026-05-26T08:14:56Z

        "properties": {
          "enabled": { "type": "boolean", "default": true, "description": "是否启用自动召回" },
          "maxResults": { "type": "number", "default": 5, "description": "召回最大结果数" },
+          "maxCharsPerMemory": { "type": "number", "default": 800, "description": "单条 L1 记忆注入的最大字符数；填 0 表示不限制" },


default should be zero

Resolved in c588558: plugin schema default is now 0.

withRiver · 2026-05-26T08:15:26Z

 | `storeBackend` | `"sqlite"` | Storage backend: `sqlite` |
 | `recall.strategy` | `"hybrid"` | Recall strategy: `keyword` / `embedding` / `hybrid` (RRF fusion, recommended) |
 | `recall.maxResults` | `5` | Number of items returned per recall |
+| `recall.maxCharsPerMemory` | `800` | Max characters injected for one recalled L1 memory; `0` disables this guard |


default should be zero

Resolved in c588558: README default is now 0.

withRiver · 2026-05-26T08:15:43Z

 | `storeBackend` | `"sqlite"` | 存储后端：`sqlite` |
 | `recall.strategy` | `"hybrid"` | 召回策略：`keyword` / `embedding` / `hybrid`（RRF 融合，推荐） |
 | `recall.maxResults` | `5` | 每次召回条数 |
+| `recall.maxCharsPerMemory` | `800` | 单条 L1 记忆注入的最大字符数；`0` 表示不限制 |


default should be zero

Resolved in c588558: README_CN default is now 0.

withRiver · 2026-05-26T08:35:33Z

There are no existing test files under src/core/hooks/. The budget logic in normalizeBudgetLimit / truncateRecallLine is straightforward enough that the config parsing coverage in src/config.test.ts plus manual verification should suffice for this PR. Let's remove this file for now.

Resolved in c588558: removed src/core/hooks/auto-recall.test.ts from this PR.

withRiver · 2026-05-26T08:36:05Z

Remove this test file.

Resolved in c588558: removed src/config.test.ts from this PR.

withRiver · 2026-05-26T08:51:24Z

+      continue;
+    }
+
+    const separatorChars = budgeted.length > 0 ? 1 : 0;


The 1 here implicitly assumes the join separator is "\n" (line 212). Consider adding a brief comment like // "\n".length to make the coupling explicit

Resolved in c588558: introduced RECALL_LINE_SEPARATOR and use its .length for budget accounting, matching the final join separator.

withRiver · 2026-05-26T09:09:06Z

+        const totalBounded = truncateRecallLine(perMemoryBounded, remainingChars);
+        budgeted.push(totalBounded);
+        usedChars += separatorChars + totalBounded.length;
+        if (totalBounded !== perMemoryBounded) truncatedCount++;


truncatedCount can be incremented twice for the same memory — once at line 734 (per-memory truncation) and again here (total-budget truncation). This makes the debug log misleading (e.g. reporting truncated=3 when only 2 distinct lines were affected). Consider tracking truncated lines with a Set or a boolean flag per iteration instead.

Resolved in c588558: truncatedCount now uses a per-iteration flag, so one emitted memory line is counted at most once even if both per-memory and total-budget truncation apply.

withRiver · 2026-05-26T09:11:42Z

+    if (perMemoryBounded.length > remainingChars) {
+      if (remainingChars >= MIN_TRUNCATED_RECALL_LINE_CHARS) {
+        const totalBounded = truncateRecallLine(perMemoryBounded, remainingChars);
+        budgeted.push(totalBounded);
+        usedChars += separatorChars + totalBounded.length;
+        if (totalBounded !== perMemoryBounded) truncatedCount++;
+      } else {
+        droppedCount++;
+      }
+      droppedCount += lines.length - i - 1;
+      break;


The droppedCount accumulation is split across the if/else branch and the line after it, making the logic harder to follow. Suggestion:

if (perMemoryBounded.length > remainingChars) { const canFit = remainingChars >= MIN_TRUNCATED_RECALL_LINE_CHARS; if (canFit) { budgeted.push(truncateRecallLine(perMemoryBounded, remainingChars)); } droppedCount += lines.length - i - (canFit ? 1 : 0); break; }

Single accumulation point, same semantics, easier to reason about.

Resolved in c588558: collapsed dropped-count handling to a single accumulation point in the total-budget overflow branch.

withRiver · 2026-05-26T12:28:38Z

LGTM. Thanks for your contribution!

By the way, there’s another issue #3 related to L1 recall. If you’re interested, feel free to take a look.

YOMXXX added 2 commits May 21, 2026 22:35

feat(recall): cap injected memory context

40eafcd

Signed-off-by: 李冠辰 <liguanchen@xiaomi.com>

fix(package): ship built runtime only

f7ee6f1

Signed-off-by: 李冠辰 <liguanchen@xiaomi.com>

This was referenced May 24, 2026

feat: L1 extraction quality improvements — reduce LLM dependency #83

Open

fix: remove duplicate system prompt in CleanContextRunner #84

Open

fix(plugin): add fallback for runtime.state.resolveStateDir on OpenClaw >=2026.5.x #79

Closed

withRiver reviewed May 25, 2026

View reviewed changes

chore(package): keep manifest unchanged

8031ec0

withRiver reviewed May 26, 2026

View reviewed changes

fix(recall): address budget review feedback

c588558

withRiver merged commit 1bdcf28 into Tencent:main May 26, 2026
5 checks passed

Conversation

YOMXXX commented May 21, 2026 • edited by withRiver Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test Plan

Notes

Uh oh!

Maxwell-Code07 commented May 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

YOMXXX commented May 25, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

withRiver commented May 26, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

YOMXXX commented May 21, 2026 •

edited by withRiver

Loading

Maxwell-Code07 commented May 22, 2026 •

edited

Loading