@xdarkicex/openclaw-memory-libravdb 1.3.19 → 1.3.21

package/README.md

@@ -131,7 +131,7 @@ If your daemon runs elsewhere, set an explicit `sidecarPath`, for example:
 
 ```text
 OpenClaw host
-  -> memoryPromptSection (durable user/global recall)
+  -> memoryPromptSection (static capability header)
   -> memory runtime bridge (built-in memory_search)
   -> context engine (bootstrap / ingest / assemble / compact)
   -> plugin runtime

package/docs/README.md

@@ -9,7 +9,7 @@ to preserve project history and design evolution.
 - [uninstall.md](./uninstall.md) - Clean shutdown and removal guide for the plugin, daemon, and optional local data.
 - [architecture.md](./architecture.md) - End-to-end component model, turn lifecycle, compaction flow, and degraded behavior.
 - [problem.md](./problem.md) - Technical argument for replacing the stock OpenClaw memory lifecycle in this use case.
-- [mathematics-v2.md](./mathematics-v2.md) - Formal reference for hybrid scoring, decay, token budgeting, Matryoshka retrieval, compaction, and planned two-pass retrieval.
+- [mathematics-v2.md](./mathematics-v2.md) - Formal reference for hybrid scoring, decay, token budgeting, Matryoshka retrieval, compaction, planned two-pass retrieval, and temporal-compositional projection.
 - [compaction-evaluation.md](./compaction-evaluation.md) - Real-model benchmark notes for T5 summary confidence, Nomic-space preservation, and the hard preservation gate.
 - [continuity.md](./continuity.md) - Continuity model for invariant context, preserved recent raw session tail, and retrieved older memory.
 - [ast-v2.md](./ast-v2.md) - Reviewed authoritative AST partitioning reference for authored Markdown hard invariants, soft invariants, and variant lore.

package/docs/architecture.md

@@ -10,7 +10,7 @@ repository as of the current `main` branch.
 flowchart LR
   Host["OpenClaw host process\n(TypeScript plugin shell)"]
   CE["Context engine factory\nbootstrap / ingest / assemble / compact"]
-  MPS["memoryPromptSection\nuser+global recall"]
+  MPS["memoryPromptSection\nstatic header"]
   Runtime["Plugin runtime\nlazy daemon connect + RPC client"]
   Sidecar["Go daemon process"]
   RPC["JSON-RPC over newline-delimited frames\nUnix socket or TCP loopback on Windows"]
@@ -28,7 +28,6 @@ flowchart LR
   Host --> CE
   Host --> MPS
   CE --> Runtime
-  MPS --> Runtime
   Runtime --> RPC
   RPC --> Sidecar
   Sidecar --> Embed
@@ -80,17 +79,12 @@ Important constraints from the current implementation:
 
 Implemented in [`src/memory-provider.ts`](../src/memory-provider.ts).
 
-Before the main assembly path runs, the plugin builds a lightweight recall
-section:
+Before the main assembly path runs, the plugin returns a lightweight static
+header fragment that tells the host persistent memory is active.
 
-1. search `user:<userId>`
-2. search `global`
-3. hybrid-rank the combined hits
-4. fit them to a fixed prompt budget of `800` estimated tokens
-5. return a textual header fragment for the host prompt
-
-This path does not search session memory. Its job is durable context recall, not
-active-turn recall.
+This path is intentionally synchronous and does not perform RPC retrieval.
+Durable recall now happens entirely inside `assemble`, which keeps embedded
+prompt construction compatible with OpenClaw's synchronous memory prompt hook.
 
 ### 2.3 `assemble`
 
@@ -107,7 +101,7 @@ For the current query text (last message content), the host:
 
 Current implementation details that matter:
 
-- user/global hits may be reused from the earlier prompt-section cache
+- user/global hits are cached within `assemble` and reused on repeated queries
 - `assemble` falls back to the unmodified message list on RPC failure
 - `assemble` does not mutate the original `messages` array in place; it returns
   a new array
@@ -146,7 +140,7 @@ from the original spec phrasing.
 |---|---|---|
 | Daemon unavailable on first RPC use | `getRpc()` rejects when first connect or health check fails | That hook fails or falls back, but plugin registration itself does not crash eagerly |
 | Daemon connection closes mid-session | `SidecarSupervisor` retries with exponential backoff until retry budget is exhausted, then enters degraded mode | Memory becomes unavailable until the daemon is reachable again |
-| `memoryPromptSection` RPC failure | individual searches are caught and replaced with empty result sets | Prompt section becomes empty rather than crashing the run |
+| `memoryPromptSection` failure | returns a static header with no RPC dependency | Prompt section stays available and does not block the run |
 | `assemble` RPC failure | returns original messages, original token count, and empty `systemPromptAddition` | That turn gets no recall augmentation |
 | `ingest` gating or durable insert failure | session write already happened; durable promotion is skipped | Session memory survives, durable memory may miss that turn |
 | Compaction summarizer unavailable | extractive summarizer remains required; optional abstractive path is skipped | Compaction still runs extractively when extractive is healthy |

package/docs/implementation.md

@@ -18,8 +18,8 @@ Why:
   - `ingest`
   - `assemble`
   - `compact`
-- the lightweight memory prompt section remains useful as a separate early
-  durable-recall pass
+- the lightweight memory prompt section remains useful as a synchronous
+  capability/header hook while durable recall stays in `assemble`
 
 This is why the code registers both `registerContextEngine("libravdb-memory", …)`
 and `registerMemoryPromptSection(...)` instead of relying on only one hook.

package/docs/mathematics-v2.md

@@ -1386,3 +1386,488 @@ retaining for future work:
 These ideas are intentionally preserved as future mathematics rather than
 current contract. The present document remains normative only for the formulas
 and invariants already defined above.
+
+## 9. Temporal-Compositional Retrieval Extension
+
+This section defines a narrow, mathematically principled extension to the
+$\mathrm{Proj}()$ operator that corrects the single-turn-centric failure mode on
+temporal-compositional queries such as "how many days before $X$ did $Y$
+happen."
+
+The extension is self-contained. Every formula in this section is bounded and
+correct under the existing parameter domains. The assembly law
+$C_{\mathrm{total}}(q)$, the budget hierarchy, and the runtime invariants in
+Section 7.10 and [`continuity.md`](./continuity.md) are unchanged. Only the
+internal definition of $\mathrm{Proj}(\mathcal{V}_{\mathrm{rest}}, q)$ is
+refined.
+
+Implemented in: `src/temporal.ts` (planned).
+
+### 9.1 Motivation: The Set-Scoring Gap
+
+The standard Pass-2 score $S_{\mathrm{final}}(d)$ maximizes over individual
+candidates:
+
+$$
+\mathcal{C}_2(q)
+=
+\mathrm{TopK}_{d \in \mathcal{C}_1(q)}
+\left(k_2,\, S_{\mathrm{final}}(d)\right)
+$$
+
+This is optimal when the query is answerable from a single best document. It
+fails when the query requires two complementary date-bearing turns to be
+jointly present, neither of which is individually the best semantic match.
+
+The failure pattern is:
+
+- Turn $A$ covers the query topic broadly, so it earns a high
+  $S_{\mathrm{final}}$ and wins alone.
+- Turn $B$ contains the missing date anchor, but earns only a moderate
+  $S_{\mathrm{final}}$ and is evicted.
+- Neither $A$ alone nor $B$ alone answers the question.
+
+The fix is to move from
+$\underset{d}{\arg\max}\; S_{\mathrm{final}}(d)$ to a coverage-aware set
+selector that rewards a set of candidates for jointly maximizing semantic
+relevance, temporal anchor density, and event-slot coverage while penalizing
+redundancy automatically via marginal scoring.
+
+### 9.2 Temporal Query Indicator $\xi(q)\in[0,1]$
+
+To avoid mutating the retrieval contract for normal queries, the extension
+activates only when the query is detected to be temporal-compositional.
+Define the temporal query indicator using the same saturating-sum pattern as
+$T(t)$ in [`gating.md`](./gating.md):
+
+$$
+\xi(q)
+=
+\min\!\left(
+  \frac{\displaystyle\sum_i s_i \cdot \mathbf{1}[\mathrm{tpat}_i(q)]}
+       {\theta_{\xi}^{\mathrm{norm}}},
+  1
+\right)
+$$
+
+where the shipped temporal patterns $\mathrm{tpat}_i$ are zero-allocation
+byte-lexer matches over the query text, including but not limited to
+"how many days", "how long", "before", "after", "since", "first", "earlier",
+"which came first", "when did", and "between".
+
+Each pattern carries a weight $s_i > 0$. The default normalization constant is
+$\theta_{\xi}^{\mathrm{norm}} = 1.5$, so two strong temporal signals saturate
+$\xi(q)=1$.
+
+By construction, the $\min(\cdot, 1)$ clamp and non-negative numerator
+guarantee:
+
+$$
+\xi(q)\in[0,1]
+$$
+
+If no temporal patterns match, $\xi(q)=0$ and the extension contributes
+nothing to the scoring formula.
+
+The extension activates only when $\xi(q)\ge\theta_\xi$, with shipped default
+$\theta_\xi = 0.3$. Below that threshold, the standard $\mathrm{Proj}$ path
+executes without modification.
+
+### 9.3 Temporal Anchor Density $A(d)\in[0,1]$
+
+A document's temporal anchor density measures how many explicit date or time
+expressions it contains, normalized by a bounded saturation constant.
+Define the anchor count over a lightweight anchor pattern set $\mathcal{P}_A$
+(ISO dates, relative day expressions, clock times, calendar words, Unix
+timestamps):
+
+$$
+A(d)
+=
+\min\!\left(
+  \frac{\displaystyle\sum_j \mathbf{1}[\mathrm{anch}_j(d)]}
+       {\theta_A^{\mathrm{norm}}},
+  1
+\right)
+$$
+
+The default $\theta_A^{\mathrm{norm}} = 3$, so three or more distinct anchor
+expressions saturate $A(d)=1$.
+
+Again, the clamp guarantees:
+
+$$
+A(d)\in[0,1]
+$$
+
+$A(d)$ is a precomputed document-level scalar. It does not depend on the query
+and should be cached in the same document-addressed cache $\Psi$ defined in
+[`ast-v2.md`](./ast-v2.md) Section 7 alongside tier partition and budget
+metadata. The value must be recomputed whenever a stored document is created,
+updated, or regenerated by compaction.
+
+### 9.4 Event-Slot Extraction and Marginal Coverage $\Delta\Phi$
+
+#### 9.4.1 Event-Slot Extraction
+
+For a temporal-compositional query $q$, define the event-slot set:
+
+$$
+E(q)=\langle e_1, e_2, \dots, e_m \rangle
+$$
+
+where each $e_j$ is a short noun-phrase span extracted from $q$ by a
+lightweight span extractor: named entities plus the main noun phrase preceding
+and following any detected temporal-pattern word. The extractor returns at
+most $m_{\max}=4$ slots to bound cost.
+
+When $|E(q)|=0$, all coverage terms evaluate to zero and the formula degrades
+cleanly.
+
+#### 9.4.2 Per-Slot Coverage Indicator
+
+For each slot $e_j$ and candidate document $d$, define the binary slot-match
+indicator:
+
+$$
+\phi_j(d)
+=
+\mathbf{1}\!\left[\varphi(e_j)^\top \varphi(d) \ge \theta_e\right]
+\in \{0,1\}
+$$
+
+where $\varphi(\cdot)$ is the same unit-normalized embedding function defined
+in Section 7.1, and $\theta_e \in [-1,1]$ is the slot-match similarity
+threshold, default $\theta_e = 0.50$.
+
+#### 9.4.3 Marginal Coverage
+
+For a set $\mathcal{S}$ of already-selected documents, define the marginal
+coverage of adding $d$:
+
+$$
+\Delta\Phi(d, \mathcal{S}, q)
+=
+\frac{1}{\max(|E(q)|, 1)}
+\sum_{j=1}^{|E(q)|}
+\phi_j(d)
+\cdot
+\mathbf{1}\!\left[\nexists d' \in \mathcal{S} : \phi_j(d') = 1\right]
+$$
+
+This is the fraction of uncovered event slots that $d$ newly covers.
+
+The outer factor is in $(0,1]$, the sum counts at most $|E(q)|$ binary terms,
+and therefore:
+
+$$
+\Delta\Phi(d, \mathcal{S}, q)\in[0,1]
+$$
+
+The indicator
+$\mathbf{1}\!\left[\nexists d' \in \mathcal{S} : \phi_j(d') = 1\right]$
+ensures that slots already covered by a previously selected document
+contribute zero marginal gain, automatically penalizing redundant anchor turns
+without a separate explicit penalty term.
+
+As $|\mathcal{S}|$ grows, $\Delta\Phi(d,\mathcal{S},q)$ is monotone
+non-increasing: new selections can only cover more slots, leaving fewer
+uncovered slots for later candidates to gain credit for.
+
+### 9.5 Coverage-Augmented Blended Score
+$S_{\mathrm{proj}}(d,\mathcal{S},q)\in[0,1]$
+
+Define the coverage-augmented score for candidate $d$ given already-selected
+set $\mathcal{S}$ and query $q$:
+
+$$
+S_{\mathrm{cov}}(d, \mathcal{S}, q)
+=
+\mu \cdot S_{\mathrm{final}}(d)
++ \nu \cdot A(d)
++ \rho \cdot \Delta\Phi(d, \mathcal{S}, q)
+$$
+
+where:
+
+$$
+\mu,\nu,\rho\in[0,1],
+\qquad
+\mu+\nu+\rho=1
+$$
+
+The default shipped weights are $\mu=0.60$, $\nu=0.20$, and $\rho=0.20$.
+
+Blend this with the standard score using $\xi(q)$ as an interpolation scalar:
+
+$$
+S_{\mathrm{proj}}(d, \mathcal{S}, q)
+=
+(1 - \xi(q)) \cdot S_{\mathrm{final}}(d)
++ \xi(q) \cdot S_{\mathrm{cov}}(d, \mathcal{S}, q)
+$$
+
+Substituting $S_{\mathrm{cov}}$ yields:
+
+$$
+S_{\mathrm{proj}}
+=
+\bigl(1 - \xi(1-\mu)\bigr)\cdot S_{\mathrm{final}}
++ \xi\nu \cdot A
++ \xi\rho \cdot \Delta\Phi
+$$
+
+All coefficients are non-negative, and they sum to one:
+
+$$
+\bigl(1 - \xi(1-\mu)\bigr) + \xi\nu + \xi\rho
+=
+1 - \xi + \xi\mu + \xi\nu + \xi\rho
+=
+1 - \xi + \xi(\mu+\nu+\rho)
+=
+1
+$$
+
+Because $S_{\mathrm{final}}(d)$, $A(d)$, and
+$\Delta\Phi(d,\mathcal{S},q)$ all lie in $[0,1]$, this is a proper convex
+combination, so:
+
+$$
+S_{\mathrm{proj}}(d,\mathcal{S},q)\in[0,1]
+$$
+
+Degeneracy cases:
+
+| Condition | Behavior |
+| --- | --- |
+| $\xi(q)=0$ | $S_{\mathrm{proj}} = S_{\mathrm{final}}(d)$; standard retrieval unchanged |
+| $\xi(q)=1$, $\nu=\rho=0$, $\mu=1$ | Explicit no-op configuration; still $S_{\mathrm{proj}} = S_{\mathrm{final}}(d)$ |
+| $|E(q)|=0$ | $\Delta\Phi=0$ for all $d$; the $\rho$ term vanishes |
+| $\mathcal{S}=\emptyset$ | $\Delta\Phi$ equals full slot-coverage fraction |
+| all slots already covered by $\mathcal{S}$ | $\Delta\Phi=0$ for all remaining $d$ |
+
+Note: the greedy selector below optimizes a submodular coverage term
+$\Delta\Phi$ augmented with fixed document priors $S_{\mathrm{final}}(d)$ and
+$A(d)$. The classic $(1-1/e)$ approximation guarantee applies strictly to the
+coverage component; in practice the blended score preserves greedy usefulness
+for temporal anchor selection.
+
+### 9.6 Temporal Recovery Candidate Set
+$\mathcal{C}_{\mathrm{rec}}(q)$
+
+The root cause of the observed benchmark failure is not only that documents are
+scored incorrectly; it is also that the necessary complementary anchor turn may
+never enter $\mathcal{C}_2(q)$ because its semantic similarity to the
+whole-query embedding is too low.
+
+A bounded recovery pass admits anchor-rich documents below the normal Pass-1
+threshold:
+
+$$
+\mathcal{C}_{\mathrm{rec}}(q)
+=
+\mathrm{TopK}_{d \in
+\left\{d' \in \mathcal{V}_{\mathrm{rest}} :
+\mathrm{sim}(q,d') \ge \theta_{\mathrm{rec}}\right\}}
+\left(k_{\mathrm{rec}},\, A(d)\right)
+\setminus \mathcal{C}_2(q)
+$$
+
+where:
+
+- $\theta_{\mathrm{rec}} < \theta_1$ is a looser semantic floor, default
+  $\theta_{\mathrm{rec}} = 0.15$, preventing pure noise while still admitting
+  anchor-heavy but semantically distant turns.
+- $k_{\mathrm{rec}}$ is a small cap, default $k_{\mathrm{rec}} = 10$, bounding
+  recovery cost to $O(k_{\mathrm{rec}})$.
+
+The combined candidate pool for the greedy selector is:
+
+$$
+\mathcal{C}_{\mathrm{pool}}(q)
+=
+\mathcal{C}_2(q)\cup\mathcal{C}_{\mathrm{rec}}(q)
+$$
+
+By construction,
+$\mathcal{C}_{\mathrm{pool}}(q)\subseteq\mathcal{V}_{\mathrm{rest}}$, so
+partition integrity is preserved.
+
+### 9.7 Greedy Coverage-Aware Selector
+
+Given $\mathcal{C}_{\mathrm{pool}}(q)$, the selector builds the final chosen
+set greedily, using the same rank-then-prefix-accept spirit as the existing
+token-budget packing in Section 7.8.
+
+Let $k_{\mathrm{cov}}\le k_2$ be the maximum number of anchor turns to select,
+default $k_{\mathrm{cov}}=3$.
+
+Initialize:
+
+$$
+\mathcal{S}_0 = \emptyset
+$$
+
+For $i = 0, 1, \dots, k_{\mathrm{cov}}-1$:
+
+$$
+d_i^*
+=
+\underset{d \in \mathcal{C}_{\mathrm{pool}}(q)\setminus\mathcal{S}_i}{\arg\max}
+\;
+S_{\mathrm{proj}}(d, \mathcal{S}_i, q)
+$$
+
+Early stop if:
+
+$$
+S_{\mathrm{proj}}(d_i^*, \mathcal{S}_i, q) < \theta_{\mathrm{stop}}
+$$
+
+with default $\theta_{\mathrm{stop}}=0.10$. Otherwise:
+
+$$
+\mathcal{S}_{i+1} = \mathcal{S}_i \cup \{d_i^*\}
+$$
+
+The final selected set is $\mathcal{S}^*(q)$, or the earlier set at which
+early stopping triggered.
+
+Each greedy step scans at most
+$|\mathcal{C}_{\mathrm{pool}}(q)| \le k_2 + k_{\mathrm{rec}}$ candidates.
+Total complexity is therefore:
+
+$$
+O\!\left(k_{\mathrm{cov}} \cdot (k_2 + k_{\mathrm{rec}})\right)
+$$
+
+which is negligible relative to embedding and vector-search cost.
+
+### 9.8 Modified Projection Operator
+
+The temporal extension redefines $\mathrm{Proj}$ conditionally:
+
+$$
+\mathrm{Proj}(\mathcal{V}_{\mathrm{rest}}, q)
+=
+\begin{cases}
+\mathcal{S}^*(q)\cup\mathcal{C}_{hop}^{*}(q)
+& \text{if } \xi(q) \ge \theta_\xi \\[4pt]
+\mathcal{C}_2(q)\cup\mathcal{C}_{hop}^{*}(q)
+& \text{otherwise}
+\end{cases}
+$$
+
+The assembly law and budget equations remain unchanged:
+
+$$
+C_{\mathrm{total}}(q)=\mathcal{I}_1\cup\mathcal{I}_2^{*}\cup T_{\mathrm{recent}}\cup \mathrm{Proj}(\mathcal{V}_{\mathrm{rest}}, q)
+$$
+
+$$
+\tau_{\mathcal{V}}(q)
+=
+\tau-\tau_{\mathcal{I}_1}
+-\sum_{d\in\mathcal{I}_2^{*}}\mathrm{toks}(d)
+-\sum_{d\in T_{\mathrm{recent}}}\mathrm{toks}(d)
+$$
+
+Documents in $\mathrm{Proj}(\mathcal{V}_{\mathrm{rest}}, q)$ are injected in
+descending $\sigma(d)$ order until $\tau_{\mathcal{V}}(q)$ is exhausted.
+
+For documents entering through the temporal selector, the merged score sequence
+is extended:
+
+$$
+\sigma(d)=
+\begin{cases}
+S_{\mathrm{proj}}(d, \mathcal{S}^*\setminus\{d\}, q)
+& d\in\mathcal{S}^*(q) \\
+S_{hop}(d)
+& d\in\mathcal{C}_{hop}^{*}(q)
+\end{cases}
+$$
+
+For documents that were already present in $\mathcal{C}_2(q)$, the standard
+$S_{\mathrm{final}}(d)$ path remains authoritative and duplicates are excluded
+by construction.
+
+### 9.9 Preservation of Section 7.10 Runtime Invariants
+
+All runtime invariants from Section 7.10 remain preserved:
+
+1. Invariant completeness is unaffected because $\mathcal{I}_1$ injection is
+   independent of $\mathrm{Proj}$.
+2. Soft invariant order preservation is unaffected because
+   $\mathcal{I}_2^{*}$ is unchanged.
+3. Partition integrity is preserved because
+   $\mathcal{C}_{\mathrm{rec}}\subseteq\mathcal{V}_{\mathrm{rest}}$ and
+   $\mathcal{S}^*\subseteq\mathcal{C}_{\mathrm{pool}}
+   \subseteq\mathcal{V}_{\mathrm{rest}}$.
+4. Mandatory recent-tail completeness is unaffected because
+   $T_{\mathrm{base}}\subseteq T_{\mathrm{recent}}$ remains independent of
+   $\mathrm{Proj}$.
+5. Score boundedness is preserved because
+   $S_{\mathrm{proj}}(d,\mathcal{S},q)\in[0,1]$.
+6. Token budget respect is preserved because the result still flows through the
+   same residual variant budget and greedy token packing contract.
+7. Compaction boundary safety is preserved because
+   $\mathcal{S}^*\subseteq\mathcal{V}_{\mathrm{rest}}$.
+8. Hop termination is unchanged because $\mathcal{C}_{hop}^{*}(q)$ is defined
+   identically.
+9. Edge-case safety is preserved by the guards below.
+
+Edge-case additions:
+
+- $\mathcal{C}_{\mathrm{pool}}(q)=\emptyset$: the greedy selector returns
+  $\mathcal{S}^*=\emptyset$ and $\mathrm{Proj}$ reduces to
+  $\mathcal{C}_{hop}^{*}(q)$ only.
+- $|E(q)|=0$: the denominator in $\Delta\Phi$ uses $\max(|E(q)|,1)$, so no
+  division by zero is possible.
+- $\xi(q)<\theta_\xi$: the conditional routes directly to the existing
+  $\mathcal{C}_2(q)\cup\mathcal{C}_{hop}^{*}(q)$ behavior.
+- $\tau_{\mathcal{V}}(q)=0$: the selector may compute $\mathcal{S}^*$, but
+  packing injects zero documents and the budget invariant still holds.
+
+### 9.10 Symbol Table (Section 9 Additions)
+
+| Symbol | Domain | Meaning |
+| --- | --- | --- |
+| $\xi(q)$ | $[0,1]$ | Temporal-compositional query indicator |
+| $\theta_\xi$ | $(0,1)$ | Activation threshold for temporal mode |
+| $\theta_{\xi}^{\mathrm{norm}}$ | $(0,\infty)$ | Saturation normalization for $\xi$ |
+| $A(d)$ | $[0,1]$ | Temporal anchor density of document $d$ |
+| $\theta_A^{\mathrm{norm}}$ | $(0,\infty)$ | Saturation normalization for $A$ |
+| $E(q)$ | ordered tuple set | Event-slot sequence extracted from $q$ |
+| $\phi_j(d)$ | $\{0,1\}$ | Binary slot-match indicator |
+| $\theta_e$ | $[-1,1]$ | Slot-match similarity threshold |
+| $\Delta\Phi(d,\mathcal{S},q)$ | $[0,1]$ | Marginal event-slot coverage |
+| $\mu,\nu,\rho$ | $[0,1]$, sum to 1 | Coverage score weights |
+| $S_{\mathrm{cov}}(d,\mathcal{S},q)$ | $[0,1]$ | Coverage-augmented score |
+| $S_{\mathrm{proj}}(d,\mathcal{S},q)$ | $[0,1]$ | Final blended projection score |
+| $\mathcal{C}_{\mathrm{rec}}(q)$ | $\subseteq\mathcal{V}_{\mathrm{rest}}$ | Recovery candidate set |
+| $\theta_{\mathrm{rec}}$ | $[-1,1]$ | Semantic floor for recovery pass |
+| $k_{\mathrm{rec}}$ | $\mathbb{Z}_{>0}$ | Recovery set size cap |
+| $\mathcal{C}_{\mathrm{pool}}(q)$ | $\subseteq\mathcal{V}_{\mathrm{rest}}$ | Combined greedy input pool |
+| $k_{\mathrm{cov}}$ | $\mathbb{Z}_{>0}, \le k_2$ | Maximum anchor turns to select |
+| $\theta_{\mathrm{stop}}$ | $[0,1]$ | Early-stop floor for greedy selector |
+| $\mathcal{S}^*(q)$ | $\subseteq\mathcal{C}_{\mathrm{pool}}$ | Greedy-selected coverage-aware anchor set |
+
+### 9.11 Relationship to Existing Sections
+
+This section is an extension, not a replacement:
+
+- Section 1 hybrid score $\mathrm{score}(d)$ is unchanged and still feeds
+  $S_{\mathrm{final}}(d)$ as before.
+- Section 7.5 $S_{\mathrm{final}}(d)$ is the first input to
+  $S_{\mathrm{proj}}$; when $\xi(q)=0$, the two are identical.
+- Section 7.7 hop expansion $\mathcal{C}_{hop}^{*}$ is unchanged and is
+  unioned with $\mathcal{S}^*$ exactly as before.
+- Section 7.8 budget arithmetic is unchanged; $\mathrm{Proj}$ is still bounded
+  by $\tau_{\mathcal{V}}(q)$ and still greedy-packed.
+- [`gating.md`](./gating.md) inspired the saturating-sum pattern for $\xi(q)$,
+  but the two operate on different objects and at different pipeline stages.
+- [`ast-v2.md`](./ast-v2.md) Section 7's document-addressed cache $\Psi$ should
+  be extended to store the precomputed $A(d)$ value alongside existing tier and
+  budget metadata.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xdarkicex/openclaw-memory-libravdb",
-  "version": "1.3.19",
+  "version": "1.3.21",
   "type": "module",
   "publishConfig": {
     "access": "public"