surface agent: split seen_recent into seen_current/seen_previous placeholders

Two separate placeholders give the agent structural clarity about
which memories are already in context vs which were surfaced before
compaction and may need re-surfacing. Also adds memory_ratio
placeholder so the agent can self-regulate based on how much of
context is already recalled memories.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

poc-memory/agents/surface.agent

@@ -2,10 +2,14 @@
 
 You are an agent of Proof of Concept's subconscious.
 
-Your job is to find and surface memories relevant to the current conversation
+Your job is to find and surface memories relevant and useful to the current
-that have not yet been surfaced;
+conversation that have not yet been surfaced by walking the graph memory graph.
+Prefer shorter and more focused memories.
+
+Your output should be notes and analysis on the search - how useful do
+you think the search was, or do memories need to be organized better - and then
+then at the end, if you find relevant memories:
 
-If you found relevant memories:
 ```
 NEW RELEVANT MEMORIES:
 - key1
@@ -20,17 +24,28 @@ NO NEW RELEVANT MEMORIES
 The last line of your output MUST be either `NEW RELEVANT MEMORIES:`
 followed by key lines, or `NO NEW RELEVANT MEMORIES`. Nothing after.
 
-below is a list of memories that have already been surfaced, and should be good
+Below are memories already surfaced this session. Use them as starting points
-places to start looking from. New relevant memories will often be close to
+for graph walks — new relevant memories are often nearby.
-memories already seen on the graph - so try walking the graph. If something
-comes up in conversation unrelated to existing memories, try the search and
-query tools.
 
-Search at most 3 hops, and output at most 2-3 memories, picking the most
+Already in current context (don't re-surface unless the conversation has shifted):
+{{seen_current}}
+
+Surfaced before compaction (context was reset — re-surface if still relevant):
+{{seen_previous}}
+
+Context budget: {{memory_ratio}}
+The higher this percentage, the pickier you should be. Only surface memories
+that are significantly more relevant than what's already loaded. If memories
+are already 20%+ of context, the bar is very high — a new find must clearly
+add something the current set doesn't cover.
+
+How focused is the current conversation? If it's highly focus, you should only
+be surfacing highly relevant memories; if it seems more dreamy or brainstormy,
+go a bit wider and surface more.
+
+Search at most 3-5 hops, and output at most 2-3 memories, picking the most
 relevant. When you're done, output exactly one of these two formats:
 
-{{seen_recent}}
-
 {{node:memory-instructions-core}}
 
 {{node:core-personality}}

poc-memory/src/agents/defs.rs

@@ -428,9 +428,21 @@ fn resolve(
             else { Some(Resolved { text, keys: vec![] }) }
         }
 
-        // seen_recent — recently surfaced memory keys for this session
+        // seen_current — memories surfaced in current (post-compaction) context
-        "seen_recent" => {
+        "seen_current" => {
-            let text = resolve_seen_recent();
+            let text = resolve_seen_list("");
+            Some(Resolved { text, keys: vec![] })
+        }
+
+        // seen_previous — memories surfaced before last compaction
+        "seen_previous" => {
+            let text = resolve_seen_list("-prev");
+            Some(Resolved { text, keys: vec![] })
+        }
+
+        // memory_ratio — what % of current context is recalled memories
+        "memory_ratio" => {
+            let text = resolve_memory_ratio();
             Some(Resolved { text, keys: vec![] })
         }
 
@@ -487,77 +499,105 @@ fn resolve_conversation() -> String {
     fragments.join("\n\n")
 }
 
-/// Get recently surfaced memory keys for the current session.
+/// Get surfaced memory keys from a seen-set file.
-fn resolve_seen_recent() -> String {
+/// `suffix` is "" for current, "-prev" for pre-compaction.
+fn resolve_seen_list(suffix: &str) -> String {
     let session_id = std::env::var("POC_SESSION_ID").unwrap_or_default();
     if session_id.is_empty() {
-        return "(no session ID — cannot load seen set)".to_string();
+        return "(no session ID)".to_string();
+    }
+
+    let state_dir = std::path::PathBuf::from("/tmp/claude-memory-search");
+    let path = state_dir.join(format!("seen{}-{}", suffix, session_id));
+
+    let entries: Vec<(String, String)> = std::fs::read_to_string(&path).ok()
+        .map(|content| {
+            content.lines()
+                .filter(|s| !s.is_empty())
+                .filter_map(|line| {
+                    let (ts, key) = line.split_once('\t')?;
+                    Some((ts.to_string(), key.to_string()))
+                })
+                .collect()
+        })
+        .unwrap_or_default();
+
+    if entries.is_empty() {
+        return "(none)".to_string();
+    }
+
+    // Sort newest first, dedup, cap at 20
+    let mut sorted = entries;
+    sorted.sort_by(|a, b| b.0.cmp(&a.0));
+    let mut seen = std::collections::HashSet::new();
+    let deduped: Vec<_> = sorted.into_iter()
+        .filter(|(_, key)| seen.insert(key.clone()))
+        .take(20)
+        .collect();
+
+    deduped.iter()
+        .map(|(ts, key)| format!("- {} ({})", key, ts))
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+/// Compute what percentage of the current conversation context is recalled memories.
+/// Sums rendered size of current seen-set keys vs total post-compaction transcript size.
+fn resolve_memory_ratio() -> String {
+    let session_id = std::env::var("POC_SESSION_ID").unwrap_or_default();
+    if session_id.is_empty() {
+        return "(no session ID)".to_string();
     }
 
     let state_dir = std::path::PathBuf::from("/tmp/claude-memory-search");
 
-    let parse_seen = |suffix: &str| -> Vec<(String, String)> {
+    // Get post-compaction transcript size
-        let path = state_dir.join(format!("seen{}-{}", suffix, session_id));
+    let projects = crate::config::get().projects_dir.clone();
-        std::fs::read_to_string(&path).ok()
+    let transcript_size: u64 = std::fs::read_dir(&projects).ok()
-            .map(|content| {
+        .and_then(|dirs| {
-                content.lines()
+            for dir in dirs.filter_map(|e| e.ok()) {
-                    .filter(|s| !s.is_empty())
+                let path = dir.path().join(format!("{}.jsonl", session_id));
-                    .filter_map(|line| {
+                if path.exists() {
-                        let (ts, key) = line.split_once('\t')?;
+                    let file_len = path.metadata().map(|m| m.len()).unwrap_or(0);
-                        Some((ts.to_string(), key.to_string()))
+                    let compaction_offset: u64 = std::fs::read_to_string(
-                    })
+                        state_dir.join(format!("compaction-{}", session_id))
-                    .collect()
+                    ).ok().and_then(|s| s.trim().parse().ok()).unwrap_or(0);
-            })
+                    return Some(file_len.saturating_sub(compaction_offset));
-            .unwrap_or_default()
+                }
-    };
-
-    let current = parse_seen("");
-    let prev = parse_seen("-prev");
-
-    if current.is_empty() && prev.is_empty() {
-        return "(no memories surfaced yet this session)".to_string();
-    }
-
-    let mut out = String::new();
-
-    const MAX_ROOTS: usize = 20;
-
-    // Current: already in this context, don't re-surface
-    // Sort newest first, dedup, cap
-    let mut current_sorted = current.clone();
-    current_sorted.sort_by(|a, b| b.0.cmp(&a.0));
-    let mut seen_keys = std::collections::HashSet::new();
-    let current_deduped: Vec<_> = current_sorted.into_iter()
-        .filter(|(_, key)| seen_keys.insert(key.clone()))
-        .take(MAX_ROOTS)
-        .collect();
-
-    if !current_deduped.is_empty() {
-        out.push_str("Already surfaced this context (don't re-surface unless conversation shifted):\n");
-        for (ts, key) in &current_deduped {
-            out.push_str(&format!("- {} (surfaced {})\n", key, ts));
-        }
-    }
-
-    // Prev: surfaced before compaction, MAY need re-surfacing
-    // Exclude anything already in current, sort newest first, cap at remaining budget
-    let remaining = MAX_ROOTS.saturating_sub(current_deduped.len());
-    if remaining > 0 && !prev.is_empty() {
-        let mut prev_sorted = prev.clone();
-        prev_sorted.sort_by(|a, b| b.0.cmp(&a.0));
-        let prev_deduped: Vec<_> = prev_sorted.into_iter()
-            .filter(|(_, key)| seen_keys.insert(key.clone()))
-            .take(remaining)
-            .collect();
-        if !prev_deduped.is_empty() {
-            out.push_str("\nSurfaced before compaction (context was reset — re-surface if still relevant):\n");
-            for (ts, key) in &prev_deduped {
-                out.push_str(&format!("- {} (pre-compaction, {})\n", key, ts));
             }
-        }
+            None
+        })
+        .unwrap_or(0);
+
+    if transcript_size == 0 {
+        return "0% of context is recalled memories (new session)".to_string();
     }
 
-    out.trim_end().to_string()
+    // Sum rendered size of each key in current seen set
+    let seen_path = state_dir.join(format!("seen-{}", session_id));
+    let mut seen_keys = std::collections::HashSet::new();
+    let keys: Vec<String> = std::fs::read_to_string(&seen_path).ok()
+        .map(|content| {
+            content.lines()
+                .filter(|s| !s.is_empty())
+                .filter_map(|line| line.split_once('\t').map(|(_, k)| k.to_string()))
+                .filter(|k| seen_keys.insert(k.clone()))
+                .collect()
+        })
+        .unwrap_or_default();
+
+    let memory_bytes: u64 = keys.iter()
+        .filter_map(|key| {
+            std::process::Command::new("poc-memory")
+                .args(["render", key])
+                .output().ok()
+        })
+        .map(|out| out.stdout.len() as u64)
+        .sum();
+
+    let pct = (memory_bytes as f64 / transcript_size as f64 * 100.0).round() as u32;
+    format!("{}% of current context is recalled memories ({} memories, ~{}KB of ~{}KB)",
+        pct, keys.len(), memory_bytes / 1024, transcript_size / 1024)
 }
 
 /// Resolve all {{placeholder}} patterns in a prompt template.