OptMem/memo

#!/usr/bin/env python3
"""OptMem: a permanent, append-only memory for AI agents.

  memo wake [part [T]]     read your memory. Run first, every session.
  memo note "..."          record one memory: one line, at most 280 chars.
  memo sleep [id "..."]    do the pending compressions.
  memo recall <regex>      search every memory ever recorded.
  memo forget <lo>-<hi>    drop a bad summary; sleep rebuilds it.
  memo import <file>       bulk-load dated memories (bootstrap only).

Everything lives in $MEMORY_DIR. See README.md.
"""

import datetime
import fcntl
import os
import re
import sys

sys.path.insert(0, os.path.dirname(os.path.realpath(__file__)))
from blocks import cover  # noqa: E402

ENTRY_CHARS = 280
WAKE_LINES = 208  # ~16k tokens of dense text, in 3 parts
RAW_MAX = 16  # blocks up to this many memories compress from the raw log

# Every harness truncates a command that prints too much, and each drops a
# different piece: Claude Code cuts the middle at 30,000 chars, pi cuts the
# head at 50 KB, Codex budgets 10,000 tokens. So the memory is handed over in
# parts that fit all of them. These are transport limits, not memory limits.
PART_CHARS = 20000
PART_LINES = 500


# Records are FIXED WIDTH, so a memory or a block is found by seeking to its
# offset -- no scanning, no index file to keep in sync. Position IS identity:
# memory i lives at i*LOG_REC of LOG.txt, and block [k*s,(k+1)*s) lives at
# k*TREE_REC of TREE/<s>. Padding costs ~2x on disk and buys O(1) everywhere.
LOG_REC = 320
TREE_REC = 288


# ---------------------------------------------------------------- store

def store():
    d = os.environ.get("MEMORY_DIR")
    if not d:
        die("MEMORY_DIR is not set. Example: export MEMORY_DIR=~/memory")
    d = os.path.expanduser(d)
    # The directory is never created here. A typo in MEMORY_DIR would then
    # open an empty store, and the agent would wake with no past and start
    # writing a second identity. Making the directory IS creating the
    # identity, and that is a deliberate act: mkdir.
    if not os.path.isdir(d):
        die("MEMORY_DIR=%s does not exist.\nIf this is a new identity, run: "
            "mkdir -p %s" % (d, d))
    os.makedirs(os.path.join(d, "TREE"), exist_ok=True)
    p = os.path.join(d, "LOG.txt")
    if not os.path.exists(p):
        open(p, "a").close()
    return d


def config(d):
    """Optional overrides in $MEMORY_DIR/config. The file is never written:
    a store that kept its own copy of the defaults would freeze them, and
    updating the tool would stop changing how it behaves."""
    global ENTRY_CHARS, WAKE_LINES, PART_CHARS, PART_LINES
    p = os.path.join(d, "config")
    if not os.path.exists(p):
        return
    for line in open(p):
        line = line.split("#")[0].strip()
        if "=" not in line:
            continue
        k, v = (s.strip() for s in line.split("=", 1))
        if k == "ENTRY_CHARS":
            ENTRY_CHARS = int(v)
        elif k == "WAKE_LINES":
            WAKE_LINES = int(v)
        elif k == "PART_CHARS":
            PART_CHARS = int(v)
        elif k == "PART_LINES":
            PART_LINES = int(v)
    if ENTRY_CHARS > min(TREE_REC - 8, LOG_REC - 40):
        die("config: ENTRY_CHARS=%d does not fit the %d/%d-byte records."
            % (ENTRY_CHARS, LOG_REC, TREE_REC))


def log_path(d):
    return os.path.join(d, "LOG.txt")


def tree_path(d, size):
    return os.path.join(d, "TREE", str(size))


def count(path, rec):
    try:
        return os.path.getsize(path) // rec
    except OSError:
        return 0


def log_len(d):
    return count(log_path(d), LOG_REC)


def repair(path, rec):
    """Drop a partial trailing record left by a crash. It was never
    acknowledged. Without this the next append lands at a wrong offset and
    every later record is misaligned. Callers hold the lock."""
    try:
        n = os.path.getsize(path)
    except OSError:
        return
    if n % rec:
        with open(path, "r+b") as f:
            f.truncate(n - n % rec)


def parse(line):
    head, _, rest = line.partition(" ")
    date, _, text = rest.partition(" ")
    return int(head[1:]), date, text


def log_get(d, i):
    """(id, date, text) of memory i, in one seek."""
    with open(log_path(d), "rb") as f:
        f.seek(i * LOG_REC)
        return parse(f.read(LOG_REC).decode().rstrip())


def log_slice(d, lo, hi):
    """Memories [lo,hi) in one read. Records are sliced as BYTES and decoded
    one by one -- slicing decoded text would shift every boundary after the
    first multi-byte character."""
    with open(log_path(d), "rb") as f:
        f.seek(lo * LOG_REC)
        buf = f.read((hi - lo) * LOG_REC)
    return [parse(buf[i * LOG_REC:(i + 1) * LOG_REC].decode().rstrip())
            for i in range(hi - lo)]


def tree_get(d, lo, hi):
    """The summary of block [lo,hi), in one seek. None if not built yet."""
    size = hi - lo
    try:
        with open(tree_path(d, size), "rb") as f:
            f.seek((lo // size) * TREE_REC)
            rec = f.read(TREE_REC)
    except OSError:
        return None
    return rec.decode().rstrip() or None


def pad(text, rec):
    b = text.encode()
    if len(b) > rec - 1:
        die("Too long: %d bytes. The record holds %d." % (len(b), rec - 1))
    return b + b" " * (rec - 1 - len(b)) + b"\n"


def locked(d):
    lock = open(os.path.join(d, ".lock"), "w")
    fcntl.flock(lock, fcntl.LOCK_EX)
    return lock


def log_append(d, items):
    """Append memories, items = [(date, text)]. The only way LOG.txt ever
    changes. Ids are assigned INSIDE the lock: two sessions noting at the same
    moment must not be handed the same id. Returns the first id used."""
    lock = locked(d)
    try:
        repair(log_path(d), LOG_REC)
        base = log_len(d)
        with open(log_path(d), "ab") as f:
            for k, (date, text) in enumerate(items):
                f.write(pad("#%d %s %s" % (base + k, date, text), LOG_REC))
            f.flush()
            os.fsync(f.fileno())
        return base
    finally:
        lock.close()


def tree_put(d, lo, hi, text):
    """Write block [lo,hi). Blocks are built in order, so this only ever
    appends one record to one level file."""
    size = hi - lo
    lock = locked(d)
    try:
        p = tree_path(d, size)
        repair(p, TREE_REC)
        if count(p, TREE_REC) != lo // size:
            return False
        with open(p, "ab") as f:
            f.write(pad(text, TREE_REC))
            f.flush()
            os.fsync(f.fileno())
        return True
    finally:
        lock.close()


def tree_drop(d, lo, hi):
    """Forget block [lo,hi) and every block built from it, by truncating each
    level back to that point. Later blocks at those levels go too and are
    rebuilt; the log is never touched, so nothing is lost."""
    gone, size = [], hi - lo
    lock = locked(d)
    try:
        while size <= log_len(d):
            p, k = tree_path(d, size), lo // size
            n = count(p, TREE_REC)
            if n > k:
                gone += [(i * size, (i + 1) * size) for i in range(k, n)]
                with open(p, "r+b") as f:
                    f.truncate(k * TREE_REC)
            size *= 2
        return gone
    finally:
        lock.close()


def die(msg):
    print(msg, file=sys.stderr)
    sys.exit(1)


def plural(n, word):
    if n == 1:
        return "1 " + word
    if word.endswith("y"):
        word = word[:-1] + "ie"
    elif word.endswith(("s", "h", "x")):
        word += "e"
    return "%d %ss" % (n, word)


def check(text):
    text = text.strip()
    if not text:
        die("Empty. A memory is one line of text.")
    if "\n" in text or "\r" in text:
        die("%d lines. A memory is one line: merge them, or note them "
            "separately." % (text.count("\n") + 1))
    n = len(text.encode())
    if n > ENTRY_CHARS:
        die("Too long: %d bytes, limit %d. Accented characters cost 2 bytes. "
            "Compress it further." % (n, ENTRY_CHARS))
    return text


# ---------------------------------------------------------------- naps

def pending(d, T, limit=None):
    """Blocks that can be built and have not been, smallest first. Each level
    file holds a dense prefix, so its length says exactly how far that level
    got: this costs one stat per level, never a scan."""
    todo, size = [], 2
    while size <= T:
        have = count(tree_path(d, size), TREE_REC)
        for k in range(have, T // size):
            todo.append((k * size, (k + 1) * size))
            if limit and len(todo) >= limit:
                return todo
        size *= 2
    return todo


def pending_count(d, T):
    """How many blocks pending() would list, without listing them. A level can
    hold MORE blocks than T needs -- T is a snapshot, and memories keep
    arriving while an agent reads -- so each level is clamped at zero."""
    n, size = 0, 2
    while size <= T:
        n += max(0, T // size - count(tree_path(d, size), TREE_REC))
        size *= 2
    return n


def nap_prompt(d, lo, hi, left):
    if hi - lo <= RAW_MAX:
        body = "\n".join("  #%d %s %s" % e for e in log_slice(d, lo, hi))
    else:
        mid, halves = (lo + hi) // 2, []
        for a, b in ((lo, mid), (mid, hi)):
            s = tree_get(d, a, b)
            if s is None:
                die("Summary %d-%d is missing. Run: memo sleep" % (a, b - 1))
            halves.append("  #%d-%d %s" % (a, b - 1, s))
        body = "\n".join(halves)
    tail = "" if not left else "\n%s after this one." % (
        "1 compression remains" if left == 1 else
        "%d compressions remain" % left)
    return ("Compress memories #%d-%d into one line of at most %d characters.\n"
            "Keep every name, number, date, decision and outcome.\n"
            "Drop wording, not facts. Invent nothing.\n\n"
            "%s\n%s\n"
            "Run: memo sleep %d-%d \"<your line>\""
            % (lo, hi - 1, ENTRY_CHARS, body, tail, lo, hi - 1))


def next_nap(d, T):
    todo = pending(d, T, limit=1)
    if not todo:
        return None
    lo, hi = todo[0]
    return nap_prompt(d, lo, hi, pending_count(d, T) - 1)


# ---------------------------------------------------------------- commands

def paginate(lines):
    """Split the document into parts that survive any harness's output cap."""
    parts, cur, size = [], [], 0
    for line in lines:
        n = len(line.encode()) + 1
        if cur and (len(cur) >= PART_LINES or size + n > PART_CHARS):
            parts.append(cur)
            cur, size = [], 0
        cur.append(line)
        size += n
    if cur:
        parts.append(cur)
    return parts


def cmd_wake(d, args):
    now = log_len(d)
    k, T = 1, now
    if args:
        if len(args) > 2 or not all(a.isdigit() for a in args):
            die("usage: memo wake [part [T]]")
        k = int(args[0])
        if len(args) == 2:
            T = int(args[1])
            if T > now:
                die("T=%d, but the memory holds %s. Run: memo wake"
                    % (T, plural(now, "entry")))
    # A part is rendered as of T, so a note landing between two parts cannot
    # shift a boundary and drop a line.
    n = pending_count(d, T)
    if n:
        print("Cannot wake: %s pending. Do %s, then run memo wake again.\n"
              % (plural(n, "compression"), "it" if n == 1 else "them"))
        print(next_nap(d, T))
        sys.exit(1)
    if not T:
        print("No memories yet. Record the first with: memo note \"<one line>\"")
        print("You are awake.")
        return
    lines = []
    for lo, hi in cover(T, WAKE_LINES):
        if hi - lo == 1:
            lines.append("#%d %s %s" % log_get(d, lo))
        else:
            s = tree_get(d, lo, hi)
            if s is None:
                die("Summary %d-%d is missing. Run: memo sleep" % (lo, hi - 1))
            lines.append("#%d-%d %s" % (lo, hi - 1, s))
    parts = paginate(lines)
    if not 1 <= k <= len(parts):
        die("No part %d: the memory has %s. Run: memo wake"
            % (k, plural(len(parts), "part")))
    if len(parts) > 1:
        print("Your memory, part %d of %d, oldest first." % (k, len(parts)))
    print("\n".join(parts[k - 1]))
    if k < len(parts):
        print("Run: memo wake %d %d" % (k + 1, T))
    else:
        # always, even for a one-part memory: the contract an agent is given
        # is "run parts until one says awake", so it must always arrive
        print("You are awake.")


def cmd_note(d, args):
    if len(args) != 1:
        die("usage: memo note \"<one line, at most %d chars>\"" % ENTRY_CHARS)
    text = check(args[0])
    i = log_append(d, [(datetime.date.today().isoformat(), text)])
    print("Saved as #%d." % i)
    nap = next_nap(d, i + 1)
    if nap:
        print("\n" + nap)


def cmd_sleep(d, args):
    T, said = log_len(d), False
    if args:
        said = True
        if len(args) != 2:
            die("usage: memo sleep <lo>-<hi> \"<one line>\"")
        m = re.fullmatch(r"(\d+)-(\d+)", args[0])
        if not m:
            die("'%s' is not a block id. Copy it from the prompt." % args[0])
        lo, hi = int(m.group(1)), int(m.group(2)) + 1
        todo = pending(d, T, limit=1)
        if not todo:
            print("Nothing left to compress.")
            return
        if (lo, hi) != todo[0]:
            die("Wrong block: %s. Blocks are built in order; the next is "
                "%d-%d. Run: memo sleep"
                % (args[0], todo[0][0], todo[0][1] - 1))
        if not tree_put(d, lo, hi, check(args[1])):
            print("Another session already wrote %d-%d." % (lo, hi - 1))
        else:
            print("%d-%d saved." % (lo, hi - 1))
    nap = next_nap(d, T)
    if not nap:
        print("Nothing left to compress.")
        return
    print(("\n" if said else "") + nap)


def cmd_forget(d, args):
    """A summary can be wrong -- mistyped, or a bad compression. Drop it and
    everything built on top of it; the next sleep computes them again. The log
    is untouched, so nothing is ever actually lost."""
    if len(args) != 1:
        die("usage: memo forget <lo>-<hi>")
    m = re.fullmatch(r"(\d+)-(\d+)", args[0])
    if not m:
        die("'%s' is not a block id." % args[0])
    lo, hi = int(m.group(1)), int(m.group(2)) + 1
    size = hi - lo
    if size < 2 or size & (size - 1) or lo % size:
        die("%s is not a block. Copy the id printed by wake, like 16-31."
            % args[0])
    gone = tree_drop(d, lo, hi)
    if not gone:
        die("No summary at %s." % args[0])
    print("Forgot %s, from %d-%d up. Run: memo sleep"
          % (plural(len(gone), "summary"), gone[0][0], gone[0][1] - 1))


def cmd_recall(d, args):
    if len(args) != 1:
        die("usage: memo recall <regex>")
    try:
        pat = re.compile(args[0], re.I)
    except re.error as e:
        die("bad regex: %s" % e)
    hits = [e for e in log_slice(d, 0, log_len(d)) if pat.search(e[2])]
    if not hits:
        print("No match.")
        return
    # Newest first is what a search is usually for, and the output has to fit
    # the same cap `wake` respects.
    out, size = [], 0
    for e in reversed(hits):
        line = "#%d %s %s" % e
        size += len(line.encode()) + 1
        if size > PART_CHARS:
            break
        out.append(line)
    print("\n".join(reversed(out)))
    if len(out) < len(hits):
        print("Newest %d of %s. Narrow the regex."
              % (len(out), plural(len(hits), "match")))
    else:
        print("%s." % plural(len(hits), "match"))


def cmd_import(d, args):
    """Bulk-append historical memories: 'YYYY-MM-DD <text>' per line.
    For bootstrapping an identity from older records. Used once."""
    if len(args) != 1:
        die("usage: memo import <file>   # lines of 'YYYY-MM-DD <text>'")
    try:
        src = open(args[0]).readlines()
    except OSError as e:
        die("Cannot read %s: %s" % (args[0], e.strerror))
    last = log_get(d, log_len(d) - 1)[1] if log_len(d) else "0000-00-00"
    out = []
    for i, line in enumerate(src, 1):
        line = line.rstrip("\n")
        if not line.strip():
            continue
        date, _, text = line.partition(" ")
        if not re.fullmatch(r"\d{4}-\d{2}-\d{2}", date):
            die("line %d: expected 'YYYY-MM-DD <text>', got: %s" % (i, line))
        if date < last:
            die("line %d: date %s precedes the previous memory (%s)."
                % (i, date, last))
        text = text.strip()
        if not text or len(text.encode()) > ENTRY_CHARS:
            die("line %d: %d bytes, limit %d." % (i, len(text.encode()), ENTRY_CHARS))
        out.append((date, text))
        last = date
    if not out:
        die("%s has no memories." % args[0])
    base = log_append(d, out)
    print("Imported %s, #%d to #%d."
          % (plural(len(out), "memory"), base, base + len(out) - 1))
    n = pending_count(d, log_len(d))
    if n:
        print("%s pending. Run: memo sleep" % plural(n, "compression"))


COMMANDS = {"wake": cmd_wake, "note": cmd_note, "sleep": cmd_sleep,
            "recall": cmd_recall, "forget": cmd_forget, "import": cmd_import}


def main():
    if len(sys.argv) < 2:
        print(__doc__.strip())
        sys.exit(0)
    if sys.argv[1] not in COMMANDS:
        print("No such command: %s\n" % sys.argv[1], file=sys.stderr)
        print(__doc__.strip(), file=sys.stderr)
        sys.exit(1)
    d = store()
    config(d)
    COMMANDS[sys.argv[1]](d, sys.argv[2:])


if __name__ == "__main__":
    main()